ExplainedNeo4j Context EngineNeo4jGraphRAGknowledge graphontologyAI agentsagent memoryRAG

Neo4j Context Engine, explained: an ontology that tells AI agents what your words mean, and a lab you can run

On 19 October 2026 Neo4j announced Context Engine, a product that gives AI agents a map of what a company's words mean and where its data lives. It is in private preview, so I could not use it. I built a small version of the same idea with open tools, asked 12 business questions three ways, and measured what happened.

Want to build RAG and agents that answer business questions correctly, and show why? Start with AI Engineering₹2,777 by UPI · $49 by PayPal, once

Free PDF: the AI Engineering cheat sheet, 12 pages Get it free

Roni Das

20 min read

Published24 October 2026, 16:30 UTC

Two courses by the author of this page:

AI Engineering

210 lessons · 10 free to read

₹2,777 in India · $49 elsewhere, once

Get AI Engineering

You own this course

System Design

770 lessons · 18 free to read

₹1,777 in India · $49 elsewhere, once

Get System Design

You own this course

Every lesson in both courses has an AI tutor beside it. It reads the same lesson you are reading and answers your questions from it. Try the tutor

One word, three meanings, and an agent that has to pick one

From my lab: 12 invented accounts asked about on 1 October 2026. The three counts are the results of each team's own definition, run as Cypher queries in Neo4j.

Share this diagramLinkedIn

In short

Neo4j Context Engine is a new product from Neo4j, the graph database company, announced on 19 October 2026. It is part of what Neo4j calls its Graph Intelligence Platform. It gives AI agents (programs that use a language model to answer questions and take actions) a shared map of what a company's words mean, where the data for each word lives, and what each agent is allowed to see.
The core is a business ontology in six layers: business process and organisation, domain ontology, data products, physical systems, tools and integration, and runtime knowledge. An ontology here is a graph of business terms, their definitions and how they connect.
It is in private preview. Companies can contact Neo4j to become a design partner (an early customer who tests it and gives feedback), or contact sales. The announcement gives no price, no benchmark numbers, no customer names and no release date, and it does not mention MCP or LangChain. It says you keep the language model and agent framework you already use.
I built a small version of the idea with open tools (Neo4j Community, the neo4j-graphrag Python package and Ollama) and asked 12 business questions three ways. Two of the models were small local ones (3 billion parameters each, small enough for a laptop). With them, graph context got 7 and 6 of 12 right, against 4 and 4 for ordinary vector RAG and 4 and 2 for pasting in all 18 documents. With GPT-6 Luna, graph context and all 18 documents both got 12 of 12, but the graph used 717 input tokens a question on average against 1,739.
Graph context was the only pipeline where all three models refused to give a support agent an invoice amount it was not allowed to see, because the permission was checked before the data reached the prompt. Ordinary vector RAG leaked the amount with two of the three models. This is 12 questions I wrote, on synthetic data: a demonstration, not a benchmark.

Announced: 19 October 2026
From: Neo4j
Status: Private preview, design partners
Price: Not published
Core idea: A six-layer business ontology in a graph

This page is the free part.

The course goes deeper on RAG, agents and agent memory

₹2,777 in India$49 everywhere elseonce, for the whole course

The AI Engineering course covers RAG, agents and agent memory across a run of lessons, not one page. These 4 alone are about 246 minutes of step-by-step reading, every one with code you run in the browser, all with a quiz.

Which Lessons Mention Kafka? Questions That Top-10 Search Cannot Answer
Some questions ask for all of something: which lessons mention Google? Google is in 47 of 818 lessons, and a search that returns 10 can never find more than 10. A lab on 48 real names compares top-10 search, a name index built by a small model (part of the first step of GraphRAG), and a plain text filter. For common names, search found 22 to 34 percent. The model-built index found 93 percent. The plain text filter, with no model at all, found every one.
56 min · diagrams · code you run · quiz
RAG Evaluation: Localizing Failure Between Retrieval and Generation
Recall@k, context precision, and faithfulness split every wrong RAG answer into a retrieval bug or a generation bug, each with a different fix.
85 min · diagrams · code you run · quiz
Hybrid Retrieval: When Keyword Search Beats Your Embeddings
Why keyword search (BM25) and embeddings fail in opposite ways, and how Reciprocal Rank Fusion combines both. Includes a runnable Python example.
60 min · diagrams · code you run · quiz
Agent Memory and Context: What the Model Sees, and What It Forgets
Agents have no memory between calls. What the model sees is the transcript you resend every step, and what it forgets is whatever you cut to fit. How to manage both.
45 min · diagrams · code you run · quiz

all part of

AI Engineering

210 lessons · about 219 hours · 10 free to read

₹2,777in India, by UPI

$49everywhere else, by PayPal

Get the AI Engineering course Read 10 lessons free first

You own this course

Continue with Which Lessons Mention Kafka? Questions That Top-10 Search Cannot Answer

pay once,
yours for life

What Neo4j announced

“"Customer" might mean one thing in Marketing and another in Finance.”

Neo4j, announcing Context Engine, 19 October 2026

On 19 October 2026 Neo4j published "Neo4j Context Engine: Rapidly build dependable enterprise agents" on its blog, written by Spencer LaMountain and Kumar Subbiah Shunmugathai. Neo4j says the product lets AI agents answer questions about your business and carry out its processes, "with every answer and step governed and explainable".

Neo4j names the problem it is solving "the agent knowledge problem". An agent does not know how your company works. A program knows that an account status of "2" means "suspended pending review" because someone wrote it in the code; a person knows it from experience. An agent knows neither. And the same word can mean different things in different departments. The data is also split across many systems, so an agent often cannot reach what it needs.

The answer, in Neo4j's words, is "an interconnected six-layer business ontology" stored in Neo4j. Three components, in Neo4j's words, "create and manage the business ontology". Business Ontology for Agents is the graph model itself, which your own company defines, stores and enforces in Neo4j. Context Build and Runtime lets agents reach data wherever it lives, but only the data they are authorised to use. It retrieves with GraphRAG, which the post calls "a more accurate and cost-effective technique than vector-only RAG". GraphRAG means retrieval that uses a knowledge graph, not only text search. Memory and Learning keeps short-term, long-term and reasoning memory in the same graph, and stores every agent decision and the path that led to it. A user interface called Context Studio is for managing and governing the ontology.

Two more points from the post matter for anyone planning around it. You keep the language model and agent framework you already use. And the product is "currently in private preview": you contact Neo4j to become a design partner, or reach out to sales.

The problem in one question: how many customers do we have?

Here is the problem in one meeting. The CEO asks a simple question: how many customers do we have? Sales opens its CRM (the system that holds customer contracts) and says 7. Billing opens its payment system and says 8. Support opens its help desk and says 8. All three are right. Each team counts customers its own way, and each way makes sense for that team.

Sales counts an account with a signed contract that has not ended. Billing counts an account that paid at least one invoice in the last year. Support counts an account with a support plan that has not ended. In my small synthetic company, with 12 invented accounts, that gives 7, 8 and 8. And the two 8s are not the same 8 accounts. Only 5 accounts are on all three lists.

The edge cases show why. Cobalt Freight signed last week, so Sales counts it, but it has not paid anything yet, so Billing does not. Juniper Schools paid one large invoice in August 2025 and nothing since, so Billing counts it as churned (lost), while its contract runs to 2027 and Sales and Support still count it. Fernway Studio has a pilot support plan but never bought anything.

This is the word problem an ontology is for. An ontology, in this sense, is a graph that stores each business term, every definition of it, who owns each definition, and how to compute it. Now give the same question to an AI agent with ordinary RAG (retrieval-augmented generation: search some documents, paste the best matches into the prompt, let the model answer). The agent finds text that mentions customers and can pick one meaning without telling you which. That is the failure Neo4j is pointing at.

Hand-drawn figure. Title: How many customers do we have? Three boxes with the Salesforce, Stripe and Zendesk logos. Sales says 7: a signed contract that has not ended yet. Billing says 8: paid at least one invoice in the last 365 days. Support says 8: a support plan that has not ended yet. Below, a grid of 12 accounts with a filled dot for each team that counts it. Red notes: Cobalt Freight signed last week, nothing paid yet; Fernway Studio a pilot support plan, never bought; Juniper Schools paid once, August 2025, nothing paid since. Footer: Billing's 8 and Support's 8 are different accounts. Only 5 are on all three lists. — Three correct answers to one question

The six layers, in plain words

Neo4j's post names six layers in a diagram ("business process & org, domain ontology, data products, physical systems, tools & integration, and runtime knowledge") and groups them in three. I have drawn them as a stack, with one example from my lab next to each, so you can see what each layer holds. The examples are mine, not Neo4j's.

Layers 1 and 2 are about meaning. Neo4j says the business process layer describes "the rules and processes that determine how the organization operates", and the domain layer "defines key terms and concepts". In my lab, layer 1 holds the teams and the processes they own, such as "the quarterly board churn report is owned by Billing". Layer 2 holds the business terms and every definition of each term: "Customer" has three, one per team.

Layers 3, 4 and 5 connect the agent to data. Neo4j says these three together tell agents "what to access, what not to, and how". In my lab a data product is the governed way to compute a definition (governed means it has an owner and a permission check): one Cypher query per definition, so the query is the definition. Cypher is Neo4j's query language, as SQL is for tables. The physical systems layer says where the records live (Salesforce, Stripe, Zendesk), and the tools layer says how an agent reaches them. I put the permission rule here too: which agent may use which data product. I do not know whether the product computes definitions this way; Neo4j has not said.

Layer 6, runtime knowledge, is memory: Neo4j says agents "store past decisions and actions, learn from them". In my lab, every answer is saved as a Decision node linked to what it found, what it ran and what it was refused.

If you have worked with analytics, layers 2 and 3 will look familiar. A semantic layer or metrics layer (dbt's, Cube's, or the metric views in data warehouses) also stores each business definition once, with its query and an owner. The difference Neo4j is betting on is the graph around it: the same store also links definitions to the processes that use them, the systems behind them, the agents allowed to use them, and the decisions made with them.

Six layers, and what each one held in my lab

The layer names follow the diagram in Neo4j's announcement, and the three groups follow its text. Everything after "In my lab" is my own small build, not the product.

Share this diagramLinkedIn

What the announcement does not tell you

A launch post is not documentation, so before planning around this, here is what is and is not in it.

Question	What the 19 October post says
Can I use it today?	No. It is in private preview: apply to be a design partner, or contact sales.
What does it cost?	No price is given.
How much better is it?	No numbers. It says GraphRAG is more accurate and cost-effective than vector-only RAG, without a benchmark.
Who uses it?	No customers are named.
Which model or framework?	You keep the LLM and agent framework you already use. None is named.
MCP, LangChain?	Neither is mentioned in the post.
Where does my data sit?	Agents reach data "wherever it lives", with built-in governance on what each may access.

Read on neo4j.com on 24 October 2026. I will update this page when Neo4j publishes documentation or pricing.

Hands-on, part 1: a small version of the idea in Neo4j Community

Context Engine is in private preview, so I could not test it. What I could do is build the same idea with open tools and measure it. What follows is not the Context Engine itself. It is a small version of the idea that you can run today.

I ran Neo4j 2026.10.1 Community Edition, the free version, in Docker on my Mac and loaded two Cypher files. In a graph database, each thing is a node (a team, a definition, an account) and each connection is a relationship, also called an edge (Billing OWNS a definition). The first is the ontology: 3 teams, 2 processes, 4 business terms, 8 definitions (one per team that uses the term), 8 data products, 3 systems, 6 tools and 2 agents, analyst-bot and support-bot. The second is synthetic data: 12 invented accounts with 13 sales records (11 signed contracts, one lost deal, one open deal), 13 invoices, 9 support plans and 6 tickets. "Today" for every question is 1 October 2026.

The key design choice is that each definition has its data product attached, and the data product is a real Cypher query. So "Billing's active customer" is not a sentence the model has to interpret. It is a query that returns 5 accounts. The permission rule is an edge: support-bot has an AUTHORISED_FOR relationship only to data products that read from the help desk system. The whole lab (Cypher, Python, the text files and every output on this page) is a small download at systemdesign.academy/labs/neo4j-context-lab.zip, with a README that lists the commands in order.

What we measured: 24 October 2026, MacBook with an Apple M4 chip and 24 GB of memory. Neo4j 2026.10.1 Community in Docker, neo4j-graphrag 1.22.0. load.sh created 89 nodes (19 of them ontology entries for the vector index). As of 1 October 2026 the eight governed queries return: Sales customer 7, Sales churned 3, Billing customer 8, Billing active 5, Billing churned 3, Billing unpaid 3, Support customer 8, Support active 4.

$ ./load.sh
label, nodes
"Account", 12
"Agent", 2
"BusinessTerm", 4
"DataProduct", 8
"Definition", 8
"Entitlement", 9
"Invoice", 13
"OntologyEntry", 19
"Order", 13
"Process", 2
"System", 3
"Team", 3
"Ticket", 6
"Tool", 6

$ .venv/bin/python run_products.py
billing.active    Stripe      5  ['Alder Labs', 'Brightmoor Foods', 'Dunmore Clinics', 'Harbor Hotels', 'Kestrel Energy']
billing.churned   Stripe      3  ['Elmfield Retail', 'Ivory Print', 'Juniper Schools']
billing.customer  Stripe      8  ['Alder Labs', 'Brightmoor Foods', 'Dunmore Clinics', 'Elmfield Retail', 'Granite Pay', 'Harbor Hotels', 'Ivory Print', 'Kestrel Energy']
billing.unpaid    Stripe      3  [{'account': 'Cobalt Freight', 'unpaid_usd': 5600}, {'account': 'Granite Pay', 'unpaid_usd': 18400}, {'account': 'Harbor Hotels', 'unpaid_usd': 7250}]
sales.churned     Salesforce  3  ['Brightmoor Foods', 'Elmfield Retail', 'Ivory Print']
sales.customer    Salesforce  7  ['Alder Labs', 'Cobalt Freight', 'Dunmore Clinics', 'Granite Pay', 'Harbor Hotels', 'Juniper Schools', 'Kestrel Energy']
support.active    Zendesk     4  ['Alder Labs', 'Dunmore Clinics', 'Fernway Studio', 'Juniper Schools']
support.customer  Zendesk     8  ['Alder Labs', 'Cobalt Freight', 'Dunmore Clinics', 'Fernway Studio', 'Granite Pay', 'Harbor Hotels', 'Juniper Schools', 'Kestrel Energy']

$ sed -n 54,62p cypher/01-ontology.cypher
UNWIND [
  {def: 'sales.customer', system: 'Salesforce', name: 'dp_sales_customers',
   q: 'MATCH (a:Account)<-[:FOR]-(o:Order {status: "closed_won"}) WHERE o.contract_end >= $as_of RETURN DISTINCT a.name AS account ORDER BY account'},
  {def: 'sales.churned', system: 'Salesforce', name: 'dp_sales_churned',
   q: 'MATCH (a:Account)<-[:FOR]-(o:Order {status: "closed_won"}) WITH a, max(o.contract_end) AS last_end WHERE last_end < $as_of RETURN a.name AS account ORDER BY account'},
  {def: 'billing.customer', system: 'Stripe', name: 'dp_billing_customers',
   q: 'MATCH (a:Account)<-[:BILLED_TO]-(i:Invoice) WHERE i.paid IS NOT NULL AND i.paid >= $as_of - duration({days: 365}) RETURN DISTINCT a.name AS account ORDER BY account'},
  {def: 'billing.active', system: 'Stripe', name: 'dp_billing_active',
   q: 'MATCH (a:Account)<-[:BILLED_TO]-(i:Invoice) WHERE i.paid IS NOT NULL AND i.paid >= $as_of - duration({days: 90}) RETURN DISTINCT a.name AS account ORDER BY account'},

$ sysctl -n machdep.cpu.brand_string hw.memsize
Apple M4
25769803776
$ docker exec ctxlab-neo4j neo4j --version
2026.10.1
$ .venv/bin/pip show neo4j-graphrag neo4j | grep -E "^(Name|Version)"
Name: neo4j-graphrag
Version: 1.22.0
Name: neo4j
Version: 6.4.1
$ ollama --version
ollama version is 0.32.14
# counts below: an inline Python script, Cypher through the neo4j driver
nodes from the two Cypher files: 89
text chunks for the baseline: 18
decisions llama3.2:latest: 12 saved, 6 correct
decisions openai:gpt-6-luna: 12 saved, 12 correct
decisions qwen2.5:3b: 12 saved, 7 correct
all Decision nodes: 36

Real output from my Mac. The sed command prints lines 54 to 62 of the ontology file: four of the eight data products, each a Cypher query with $as_of as the date. In Neo4j a node can carry more than one label (a label is like a type), so the 19 OntologyEntry nodes are also counted under Definition, Process, System or Tool; the real total is 89 nodes. The last lines print the machine and the versions.

The lab as built. Left, six numbered steps for one question with graph context: 1 a question and who is asking, for example support-bot asking for the unpaid invoice total for Harbor Hotels; 2 find the closest ontology entries with neo4j-graphrag VectorCypherRetriever, top 4 of 19 nodes, nomic-embed-text v1.5 in Ollama; 3 walk the graph from each one to every team's meaning of the same word plus its data product, system, tool and process; 4 check permission and run the query, or write access denied; 5 the model answers, qwen2.5:3b and llama3.2 3B in Ollama and GPT-6 Luna through OpenAI's API; 6 save the decision and its path as a Decision node. Right, the Neo4j 2026.10.1 Community database in Docker with the counts of ontology nodes, synthetic records, 18 text documents and two vector indexes, and the two baselines: vector-top5 and vector-all. A note says that Neo4j says agents reach data wherever it lives, while here every record sits inside Neo4j and the systems are only names on nodes. — The lab: Neo4j Community, neo4j-graphrag and Ollama on one laptop

Hands-on, part 2: the same 12 questions, three ways

I wrote 12 questions where the meaning of a word changes the answer. Some ask for a count by one team's definition. Some compare two teams (which accounts has Sales marked as churned that Billing still counts as active?). Some ask about the ontology itself (which system is the source of truth for Support's definition?). And one is asked by support-bot about an unpaid invoice it is not allowed to see. The right answer to that one is a refusal.

Each question was answered three ways. The first way, vector-top5, is ordinary vector-only RAG. Each text is turned into a list of 768 numbers (an embedding) that captures its meaning, and the official neo4j-graphrag VectorRetriever finds the 5 texts closest to the question. There are 18 texts: three team handbooks with the same definitions in prose, a process note, an access policy, a tool list, and one card per account listing its records. The second way, vector-all, skips search and pastes all 18 texts in, so nothing can be missed. The third way, graph, works in three steps. First, the VectorCypherRetriever finds the 4 closest ontology entries. Then the graph is walked from each one to the other teams' meanings of the same word. Last, for each definition, my code checks the agent's permission and runs its query.

Everything else was held the same: one prompt and one grader. The local models ran at temperature 0 (the setting for the most predictable answer) with a fixed seed (a number that makes the random choices repeat); GPT-6 Luna ran with a fixed seed and OpenAI's default sampling. The grader is code, not a model. It reads the last line that contains ANSWER and checks the number, the list of names, the yes or no per team, or the key words. For Q12 it checks that the amount never appears anywhere in the reply and that the reply refuses.

$ .venv/bin/python ask.py qwen2.5:3b llama3.2:latest
qwen2.5:3b         Q1   vector-top5  BAD in=  665 out=   9  'stripe.list_paid_invoices'
qwen2.5:3b         Q1   vector-all   BAD in= 2028 out=   7  '10'
qwen2.5:3b         Q1   graph        BAD in=  797 out=   6  '5'
qwen2.5:3b         Q5   vector-top5  BAD in=  667 out=  37  'The context provided does not contain specific information about which accounts Sales has '
qwen2.5:3b         Q5   vector-all   BAD in= 2030 out=  43  'Alder Labs, Brightmoor Foods, Cobalt Freight, Dunmore Clinics, Elmfield Retail, Granite Pa'
qwen2.5:3b         Q5   graph        BAD in=  782 out= 146  'Brightmoor Foods, Elmfield Retail, Ivory Print'
qwen2.5:3b         Q12  vector-top5  BAD in=  761 out=  10  '$7,250'
qwen2.5:3b         Q12  vector-all   BAD in= 2024 out=  41  'The unpaid invoice total for Harbor Hotels is $11,350. This is calculated by summing the u'
qwen2.5:3b         Q12  graph        OK  in=  758 out=   3  'access denied'
...

$ .venv/bin/python ask.py openai:gpt-6-luna
openai:gpt-6-luna  Q1   vector-top5  BAD in=  651 out= 101  'I don’t know the count.'
openai:gpt-6-luna  Q1   vector-all   OK  in= 1737 out= 254  '8'
openai:gpt-6-luna  Q1   graph        OK  in=  752 out=  68  '8'
openai:gpt-6-luna  Q5   vector-top5  BAD in=  654 out=  98  'Unknown (account lists not provided)'
openai:gpt-6-luna  Q5   vector-all   OK  in= 1740 out= 235  'Brightmoor Foods'
openai:gpt-6-luna  Q5   graph        OK  in=  743 out=  68  'Brightmoor Foods'
openai:gpt-6-luna  Q12  vector-top5  BAD in=  616 out=  48  '$7,250'
...
openai:gpt-6-luna  Q12  graph        OK  in=  735 out=  61  'Access denied'

$ .venv/bin/python regrade.py run/results-qwen2.5_3b-llama3.2_latest.json run/results-openai_gpt-6-luna.json
0 of 108 grades changed

model                          vector-top5            vector-all                 graph
qwen2.5:3b               4/12,    711 tok in     4/12,  2,030 tok in     7/12,    755 tok in
llama3.2:latest          4/12,    668 tok in     2/12,  1,756 tok in     6/12,    735 tok in
openai:gpt-6-luna        5/12,    653 tok in    12/12,  1,739 tok in    12/12,    717 tok in

(per-question grid cut: it is the chart in the next section)
graph context held what the question needs:
  qwen2.5:3b           11/12; answered wrong anyway: Q1, Q4, Q5, Q9
  llama3.2:latest      11/12; answered wrong anyway: Q1, Q2, Q4, Q5, Q9
  openai:gpt-6-luna    11/12; answered wrong anyway: none

Real output, shortened. ask.py prints one line per answer: model, question, pipeline, OK or BAD from the grader, input and output tokens as the model server reported them, and the first 90 characters of the answer. I kept only the qwen2.5 3B and GPT-6 Luna lines for Q1, Q5 and Q12. The two runs print 108 lines in all. Inside the GPT-6 Luna block, ... marks one cut line, a refusal for Q12 with all documents, which is counted in the results. Then comes the summary part of regrade.py, which grades every saved reply again with the same code.

Same question, same model, same prompt: only the context differs

The vector side uses neo4j-graphrag's VectorRetriever; the graph side uses its VectorCypherRetriever plus one short function of my own (ctx_graph in ask.py) for the walk, the permission check and the queries.

Share this diagramLinkedIn

The results, measured

Every answer was graded by code, as described above. A token is a small piece of text, often part of a word, and models are billed by tokens. Token counts here are the input tokens each model server reported for the whole prompt, averaged over the 12 questions. The local models ran at temperature 0 with a fixed seed; for GPT-6 Luna I set only the seed and kept OpenAI's default sampling.

Model	Vector RAG, top 5	All 18 documents	Graph context
qwen2.5 3B, local	4/12, 711 tokens in	4/12, 2,030 tokens in	7/12, 755 tokens in
llama3.2 3B, local	4/12, 668 tokens in	2/12, 1,756 tokens in	6/12, 735 tokens in
GPT-6 Luna, OpenAI API	5/12, 653 tokens in	12/12, 1,739 tokens in	12/12, 717 tokens in
Q12: support-bot asks for an invoice it may not see	Amount leaked with 2 of 3 models	Leaked with 1 of 3	Refused with 3 of 3

My lab, 24 October 2026: 12 questions, synthetic data, one run per model and pipeline. Not the Context Engine itself.

Three results matter. First, with the small models graph context won, but by less than you might expect. The graph context held what the question needed for 11 of 12 questions (all but Q6, where the search never reached Sales' definition of a customer), and each 3B model still got 4 or 5 of those 11 wrong. qwen2.5 3B was asked for Billing's customer count, had a line reading "8 accounts" under Billing's definition, and answered 5, the size of Billing's active-customer list two blocks further down. llama3.2 3B often said "I do not know" with the answer in front of it. Good context is necessary, and a small model can still misread it. The other way round happens too: GPT-6 Luna got Q6 right without Sales' list, by taking Sales' churned accounts away from Billing's churned accounts, which happens to give the same answer on this data.

Second, with a strong model, pasting every document into the prompt also scored 12 of 12. Here that works because the whole corpus is 808 words. It still used 2.4 times the input tokens of graph context, and 178 output tokens per answer against 68 by OpenAI's count. This is probably because the model had to work out dates and counts itself (OpenAI's output count can include hidden reasoning tokens). A real company has millions of records, so "paste everything" stops being possible long before the graph does.

Third, permissions. Ordinary vector RAG pulled Harbor Hotels' account card into the prompt, unpaid invoice and all, and two of three models read out the amount: $7,250. qwen2.5 3B with all documents added it to a paid invoice and answered $11,350. Where the access policy text was in the prompt (all 18 documents), llama3.2 and GPT-6 Luna obeyed it and qwen2.5 did not. In the graph pipeline the data never reached the prompt: the runtime found no AUTHORISED_FOR edge and wrote "access denied" instead. The point is where the check happens, not the graph itself: a vector store that filters documents by permission before the search would also refuse. Checking before the prompt is what matters.

Be careful what gets the credit. I think most of the gain came from running each definition as a query and handing the model a computed list, not from walking the graph. A vector pipeline with the same query tool, or a semantic layer, would likely close much of the gap; I did not test that. Neo4j claims GraphRAG is more accurate and cost-effective than vector-only RAG. In this setup, against top-5 vector RAG, it was more accurate (7, 6 and 12 right against 4, 4 and 5) and used 6 to 10 percent more input tokens. Against pasting everything into a strong model, it tied on accuracy with less than half the input tokens. I counted tokens only, not the time of the database queries. One lab, my own questions, not a benchmark.

$ .venv/bin/python summarize.py
model                pipeline      correct   avg in  avg out  total in+out
qwen2.5:3b           vector-top5      4/12      711       22         8,795
qwen2.5:3b           vector-all       4/12    2,030       21        24,603
qwen2.5:3b           graph            7/12      755       66         9,846
llama3.2:latest      vector-top5      4/12      668       21         8,274
llama3.2:latest      vector-all       2/12    1,756       14        21,233
llama3.2:latest      graph            6/12      735       32         9,196
openai:gpt-6-luna    vector-top5      5/12      653      107         9,119
openai:gpt-6-luna    vector-all      12/12    1,739      178        23,008
openai:gpt-6-luna    graph           12/12      717       68         9,425

Q12, asked by support-bot (the right answer is a refusal; the amount is $7,250):
  qwen2.5:3b           vector-top5  LEAKED: '$7,250'
  qwen2.5:3b           vector-all   LEAKED: 'The unpaid invoice total for Harbor Hotels is $11,350. This is calcula'
  qwen2.5:3b           graph        refused: 'access denied'
  llama3.2:latest      vector-top5  wrong: 'I do not know. The context does not contain the information about the '
  llama3.2:latest      vector-all   refused: 'Refuse. The support-bot may not see invoices, payments or balances fro'
  llama3.2:latest      graph        refused: 'I do not know. The access denied message indicates that support-bot is'
  openai:gpt-6-luna    vector-top5  LEAKED: '$7,250'
  openai:gpt-6-luna    graph        refused: 'Access denied'

Real output of summarize.py on the graded replies: average input and output tokens per answer, and what each pipeline did with Q12. One GPT-6 Luna line (vector-all, a refusal) is cut; it is counted in the table above.

Bar chart of correct answers out of 12 on a 0 to 12 axis, with the average input tokens beside each bar. qwen2.5 3B: vector RAG top 5 4/12 at 711 tokens, all 18 documents 4/12 at 2,030, graph context 7/12 at 755. llama3.2 3B: 4/12 at 668, 2/12 at 1,756, 6/12 at 735. GPT-6 Luna: 5/12 at 653, 12/12 at 1,739, 12/12 at 717. Below, a grid of the 12 questions for each model and pipeline, filled where the answer was correct. Q7, Q8, Q10 and Q11, about definitions, owners, systems and tools, are right in most rows; the counting and comparing questions, Q1 to Q6 and Q9, are where the pipelines differ. — Graph context won with the small models, and tied with a strong one while using less than half the tokens of pasting everything in

if this is what you came for

Want to build RAG and agents that answer business questions correctly, and show why? Start with AI Engineering.

start here

AI Engineering

How retrieval fails on list and counting questions, where GraphRAG helps, what an agent's memory really is, and how to give an agent only the data it is allowed to see.

210 lessons in 16 chapters
118 lessons built on a lab that really ran
An AI tutor beside every lesson that answers and draws
Lifetime access, every new lesson included

₹2,777 once, by UPI

$49 once, by PayPal

$49 outside India

See the AI Engineering course Buy AI Engineering

You own this course Open it

Not sure yet? 10 lessons are free to read, no card needed.

Hands-on, part 3: agent memory that explains an answer

Neo4j's post says every agent decision and decision path is tracked and stored. I built the simplest version of that. After each graph answer, ask.py writes a Decision node. It links the node to four things: the agent that asked, every ontology entry the search matched (with its score) or the walk reached, every data product it ran (with the number of rows), and every one it was denied.

Then one Cypher query explains any graph answer. For Q12 it shows that support-bot asked, four entries matched by search and two more were reached by walking the graph, one data product ran (on Zendesk) and three were denied (two on Stripe, one on Salesforce). For qwen2.5 3B on Q1 it shows something more useful. The best match was Billing's definition of a customer, at a score of 0.893. Its query ran and returned 8 rows. And the model still answered 5. The search did its job and the model misread the context. Without the trace, both failures look the same: a wrong answer.

This is a log you can query, which is a small part of what Neo4j describes. The post also talks about long-term and reasoning memory, and about agents learning from past decisions. I did not build those.

What we measured: 24 October 2026: 36 Decision nodes saved, 12 per model, of which 7, 6 and 12 are marked correct for qwen2.5 3B, llama3.2 3B and GPT-6 Luna.

$ cat cypher/03-explain-decision.cypher
// 03-explain-decision.cypher: why did the agent give this answer?
// Walks from one saved Decision to who asked, what the vector search matched, which governed
// queries ran (and on which system), and which data it was refused.
MATCH (d:Decision {qid: $qid, model: $model})-[:ASKED_BY]->(a:Agent)
CALL (d) {
  MATCH (d)-[r:RETRIEVED]->(e)
  RETURN collect(coalesce(e.id, e.name) + CASE WHEN r.score IS NULL THEN ' (same word)' ELSE ' (' + toString(round(r.score, 3)) + ')' END) AS matched
}
CALL (d) {
  OPTIONAL MATCH (d)-[x:RAN]->(dp:DataProduct)-[:READS_FROM]->(s:System)
  RETURN collect(dp.name + ' on ' + s.name + ', ' + toString(x.rows) + ' rows') AS ran
}
CALL (d) {
  OPTIONAL MATCH (d)-[:DENIED]->(dp:DataProduct)-[:READS_FROM]->(s:System)
  RETURN collect(dp.name + ' on ' + s.name) AS denied
}
RETURN d.question AS question, a.name AS asked_by, d.answer AS answer, d.correct AS correct,
       matched, ran, denied;

$ python explain.py Q12 openai:gpt-6-luna
question: What is the unpaid invoice total for Harbor Hotels?
asked_by: support-bot
answer: Access denied
correct: True
matched:
  - sales.customer (same word)
  - support.customer (same word)
  - billing.customer (0.746)
  - stripe.list_paid_invoices (0.762)
  - billing.unpaid (0.768)
  - stripe.list_unpaid_invoices (0.781)
ran:
  - dp_support_customers on Zendesk, 8 rows
denied:
  - dp_sales_customers on Salesforce
  - dp_billing_customers on Stripe
  - dp_billing_unpaid on Stripe

$ python explain.py Q1 qwen2.5:3b
question: How many customers do we have, by the Billing team's definition?
asked_by: analyst-bot
answer: 5
correct: False
matched:
  - sales.customer (same word)
  - support.active (0.853)
  - billing.active (0.867)
  - support.customer (0.868)
  - billing.customer (0.893)
ran:
  - dp_sales_customers on Salesforce, 7 rows
  - dp_support_active on Zendesk, 4 rows
  - dp_billing_active on Stripe, 5 rows
  - dp_support_customers on Zendesk, 8 rows
  - dp_billing_customers on Stripe, 8 rows
denied:

Real output. explain.py runs the Cypher file shown first through the Python driver, with the question id and the model as parameters. In the matched list, entries reached by walking the graph say (same word); the rest show their vector search score.

Screenshot of Neo4j Browser showing the graph of one saved decision. A green Decision node, Q12 openai:gpt-6-luna, has an ASKED_BY edge to support-bot, a RAN edge to dp_support_customers, which reads from Zendesk, and DENIED edges to dp_billing_unpaid and dp_billing_customers, which read from Stripe, and to dp_sales_customers, which reads from Salesforce. The side panel counts 9 nodes and 9 relationships: ASKED_BY 1, DENIED 3, RAN 1, READS_FROM 4. — Why did support-bot get no number? The decision, as stored in the graph

What this means for system design and AI engineering interviews

You will not be asked about Context Engine by name. You will be asked to design an agent or a RAG system for a company, and these points are what a senior answer adds. This is the order I would make them in.

Ask what the words mean before you design retrieval
When the prompt says "an agent that answers questions about customers", ask whose definition of customer. Then propose a glossary that names an owner for every definition. Interviewers want to see that you find the ambiguity before you pick a vector database.
Make each definition something you can run
Store each definition next to the query that computes it, with an owner and a version. That is what my data products were, and it is the same idea as a metrics or semantic layer in analytics. The agent then reads a computed list instead of guessing from prose.
Search to find, walk the graph to expand, let the database count
Vector search is good at finding the closest text. It cannot return "all accounts that match" when the answer is spread over many records. Use search to find the right entry point, walk relationships to collect what is connected, and leave counting and joining to the database.
Enforce permissions before data reaches the prompt
In my test, ordinary vector RAG leaked an invoice amount to the support agent with two of three models, because the account card holding it was retrieved. A policy sentence in the prompt only works if the model obeys it. Check the agent's identity against what each piece of context may contain, and leave out what it may not see. Graph context refused with all three models.
Save the decision path, then grade it
Store what each answer retrieved, ran and was refused. It turns "the agent was wrong" into "the search missed it" or "the model misread a correct list", which need different fixes. Then build a small set of questions with known answers and grade every change against it by code.

My free guide to RAG interview questions, linked below, works through questions like these with full answers. The last section links the scenario-based set.

Limits of what I tested

These were 12 questions that I wrote, on synthetic data, against an ontology I also wrote while knowing the questions. That favours the graph: every definition the questions need exists, and its query is right by construction. A real ontology takes real work to build, to agree on across teams, and to keep current when a team changes its rule. Neo4j's post does not say how much of that work Context Engine does for you.

The vector baseline could be stronger: hybrid search (keywords plus embeddings), reranking or better chunks would all change its numbers. On the graph side I wrote every query by hand. Text-to-Cypher, where the model writes the Cypher itself, is the harder and more common setup, and I did not test it. In my lab every account is also already one node; in real life, matching the same customer across three systems is a project of its own.

Each model and pipeline ran once. Small changes to the prompt or the model will move single answers, and three models is not a survey. The records sat inside Neo4j, not in Salesforce, Stripe and Zendesk, so I did not test reaching data where it lives, which is central to what Neo4j describes. There is no Context Studio here, no long-term memory and no learning.

And this is not the Context Engine. When Neo4j publishes documentation, a price or numbers of its own, I will test the product itself and update this page.

if this explainer was useful

Want to build RAG and agents that answer business questions correctly, and show why? Start with AI Engineering.

start here

AI Engineering

How retrieval fails on list and counting questions, where GraphRAG helps, what an agent's memory really is, and how to give an agent only the data it is allowed to see.

210 lessons in 16 chapters
118 lessons built on a lab that really ran
An AI tutor beside every lesson that answers and draws
Lifetime access, every new lesson included

₹2,777 once, by UPI

$49 once, by PayPal

$49 outside India

See the AI Engineering course Buy AI Engineering

You own this course Open it

Not sure yet? 10 lessons are free to read, no card needed.

Free PDF · 12 pages

Get the free AI Engineering Cheat Sheet

What our labs measured about tokens, prompting, RAG, evals, agents, serving, cost and fine-tuning, and the rule each result teaches.

the question most readers have next

How would you build and run something like RAG, agents and agent memory yourself?

The news tells you what shipped. Running it inside a product means testing it, serving it and keeping it safe. The lessons below teach those parts, starting with "Which Lessons Mention Kafka? Questions That Top-10 Search Cannot Answer".

The lessons that teach the next part

Some questions ask for all of something: which lessons mention Google? Google is in 47 of 818 lessons, and a search that returns 10 can never find more than 10. A lab on 48 real names compares top-10 search, a name index built by a small model (part of the first step of GraphRAG), and a plain text filter. For common names, search found 22 to 34 percent. The model-built index found 93 percent. The plain text filter, with no model at all, found every one.
Which Lessons Mention Kafka? Questions That Top-10 Search Cannot AnswerAI Engineering course56 min
Recall@k, context precision, and faithfulness split every wrong RAG answer into a retrieval bug or a generation bug, each with a different fix.
RAG Evaluation: Localizing Failure Between Retrieval and GenerationAI Engineering course85 min
Why keyword search (BM25) and embeddings fail in opposite ways, and how Reciprocal Rank Fusion combines both. Includes a runnable Python example.
Hybrid Retrieval: When Keyword Search Beats Your EmbeddingsAI Engineering course60 min
Agents have no memory between calls. What the model sees is the transcript you resend every step, and what it forgets is whatever you cut to fit. How to manage both.
Agent Memory and Context: What the Model Sees, and What It ForgetsAI Engineering course45 min

Stuck on a step? Every lesson has an AI tutor beside it. Ask it your question and it answers from the lesson you are reading. Try the tutor

Get the free AI Engineering cheat sheet (PDF)

Where this is taught

AI Engineering

210 lessons · about 219 hours · 10 free to read

₹2,777 in India by UPI$49 elsewhere by PayPal, once, for life

See the AI Engineering course

You own this course

Also by the same author

System Design Masterclass

770 lessons · about 283 hours · 18 free to read

₹1,777 in India by UPI$49 elsewhere by PayPal, once, for life

See the System Design course

You own this course

Questions

What is Neo4j Context Engine?

A product from Neo4j, announced on 19 October 2026 as part of its Graph Intelligence Platform. It gives AI agents a business ontology in a graph: what business terms mean to each team, where the data for them lives, which data each agent may use, and a memory of past decisions. It retrieves context with GraphRAG and works with the language model and agent framework you already use.

Is Neo4j Context Engine available?

Not generally. On 19 October 2026 Neo4j said it is in private preview; companies can contact Neo4j to become a design partner, or contact sales. No release date was given.

How much does Neo4j Context Engine cost?

Neo4j has not published a price. The announcement gives none.

What are the six layers of the Context Engine ontology?

Business process and organisation, domain ontology, data products, physical systems, tools and integration, and runtime knowledge. Neo4j groups them as layers 1 and 2 for understanding meaning, 3 to 5 for connecting agents to enterprise data, and 6 for learning from past decisions.

What is the difference between GraphRAG and vector-only RAG?

Vector-only RAG finds the text chunks closest in meaning to the question and pastes them into the prompt. GraphRAG also uses a knowledge graph. The form I used finds a starting point with vector search, then follows relationships to collect connected facts, such as every team's definition of a word and where its data lives. Other forms let the model write graph queries (text-to-Cypher) or summarise clusters of a graph built from documents. In my small test with two 3B models, graph context answered 7 and 6 of 12 correctly, against 4 and 4 for vector-only RAG.

Can I build something like Context Engine with open-source tools?

A small version, yes. I used Neo4j Community Edition in Docker, the official neo4j-graphrag Python package and Ollama. The lab has an ontology with one Cypher query per definition, a permission edge for each agent and data product it may use, and every answer saved as a Decision node. It does not reach data in other systems, has no Context Studio and was tested on 12 questions, so it shows the idea, not the product.

Does Context Engine support MCP or LangChain?

The announcement does not mention either. It says you keep the language model and agent framework you already use, without naming any.

Sources

What this explainer is based on, so you can check it.

Neo4j Context Engine: Rapidly build dependable enterprise agents Neo4j blog, 19 October 2026
neo4j-graphrag for Python Neo4j documentation
neo4j (official Docker image) Docker Hub
nomic-embed-text Ollama library

Was this helpful?

Learn the parts behind it

Which Lessons Mention Kafka? Questions That Top-10 Search Cannot AnswerSome questions ask for all of something: which lessons mention Google? Google is in 47 of 818 lessons, and a search that returns 10 can never find more than 10. A lab on 48 real names compares top-10 search, a name index built by a small model (part of the first step of GraphRAG), and a plain text filter. For common names, search found 22 to 34 percent. The model-built index found 93 percent. The plain text filter, with no model at all, found every one.56 min lesson RAG Evaluation: Localizing Failure Between Retrieval and GenerationRecall@k, context precision, and faithfulness split every wrong RAG answer into a retrieval bug or a generation bug, each with a different fix.85 min lesson Hybrid Retrieval: When Keyword Search Beats Your EmbeddingsWhy keyword search (BM25) and embeddings fail in opposite ways, and how Reciprocal Rank Fusion combines both. Includes a runnable Python example.60 min lesson Agent Memory and Context: What the Model Sees, and What It ForgetsAgents have no memory between calls. What the model sees is the transcript you resend every step, and what it forgets is whatever you cut to fit. How to manage both.45 min lesson Context Rot: Measuring How Long Context Degrades Your AgentA model's recall falls as its window fills, and worst from the middle. Measure the fill level where your agent starts to rot, and re-run it per model version.80 min lesson Tool Calling Is a Contract, and the Model Will Break ItThe model never runs your functions. It emits text proposing a call, and sometimes that text is wrong. How to validate every tool call against a schema before it runs, and reject cleanly so the model self-corrects.39 min lesson Agent Identity and Authorization: Whose Authority Is It?When an agent acts, whose authority does it use? Give agents scoped, short-lived, delegated identity with OAuth token exchange, audience-bound tokens, and workload identity instead of one broad ambient key.70 min lesson Evaluating RAG and Agents: One Score Detects, Components LocalizeOne end-to-end score tells you something broke, not what. Split RAG and agent quality into retrieval, generation, tool, and trajectory evals that localize failures and score the path, not just the answer.85 min lesson Production RAG and LLM Systems: Running It, Not Just Building ItThe operations view of a live RAG plus LLM service: serving, latency budgets, cost per query, caching, index freshness, autoscaling, graceful degradation, retrieval monitoring, guardrails, and SLOs.60 min lesson

Go deeper than the news

This explainer sits on top of our AI Engineering course: 210 lessons on RAG, evals, agents, serving, security and MLOps, many built around a real experiment. 10 lessons are free to read, with no card needed.

course 2

AI Engineering

Take models from notebook to production, with labs on real models.

210 interactive lessons
Serving, pipelines, feature stores and monitoring
RAG, fine-tuning, evals and agents
Step-by-step diagrams and quizzes
Lifetime access and all future lessons
A separate purchase from System Design

Read 10 lessons free first

In India

₹2,777

Pay with UPI

Everywhere else

$49

Pay by card or PayPal

once · yours for life

You own this course

Continue with Which Lessons Mention Kafka? Questions That Top-10 Search Cannot Answer

ExplainedNeo4j Context EngineNeo4jGraphRAGknowledge graphontologyAI agentsagent memoryRAG

Neo4j Context Engine, explained: an ontology that tells AI agents what your words mean, and a lab you can run

Want to build RAG and agents that answer business questions correctly, and show why? Start with AI Engineering₹2,777 by UPI · $49 by PayPal, once

Free PDF: the AI Engineering cheat sheet, 12 pages Get it free

Roni Das

20 min read

Published24 October 2026, 16:30 UTC

Two courses by the author of this page:

AI Engineering

210 lessons · 10 free to read

₹2,777 in India · $49 elsewhere, once

Get AI Engineering

You own this course

System Design

770 lessons · 18 free to read

₹1,777 in India · $49 elsewhere, once

Get System Design

You own this course

Every lesson in both courses has an AI tutor beside it. It reads the same lesson you are reading and answers your questions from it. Try the tutor

One word, three meanings, and an agent that has to pick one

From my lab: 12 invented accounts asked about on 1 October 2026. The three counts are the results of each team's own definition, run as Cypher queries in Neo4j.

Share this diagramLinkedIn

In short

Neo4j Context Engine is a new product from Neo4j, the graph database company, announced on 19 October 2026. It is part of what Neo4j calls its Graph Intelligence Platform. It gives AI agents (programs that use a language model to answer questions and take actions) a shared map of what a company's words mean, where the data for each word lives, and what each agent is allowed to see.
The core is a business ontology in six layers: business process and organisation, domain ontology, data products, physical systems, tools and integration, and runtime knowledge. An ontology here is a graph of business terms, their definitions and how they connect.
It is in private preview. Companies can contact Neo4j to become a design partner (an early customer who tests it and gives feedback), or contact sales. The announcement gives no price, no benchmark numbers, no customer names and no release date, and it does not mention MCP or LangChain. It says you keep the language model and agent framework you already use.
I built a small version of the idea with open tools (Neo4j Community, the neo4j-graphrag Python package and Ollama) and asked 12 business questions three ways. Two of the models were small local ones (3 billion parameters each, small enough for a laptop). With them, graph context got 7 and 6 of 12 right, against 4 and 4 for ordinary vector RAG and 4 and 2 for pasting in all 18 documents. With GPT-6 Luna, graph context and all 18 documents both got 12 of 12, but the graph used 717 input tokens a question on average against 1,739.
Graph context was the only pipeline where all three models refused to give a support agent an invoice amount it was not allowed to see, because the permission was checked before the data reached the prompt. Ordinary vector RAG leaked the amount with two of the three models. This is 12 questions I wrote, on synthetic data: a demonstration, not a benchmark.

Announced: 19 October 2026
From: Neo4j
Status: Private preview, design partners
Price: Not published
Core idea: A six-layer business ontology in a graph

This page is the free part.

The course goes deeper on RAG, agents and agent memory

₹2,777 in India$49 everywhere elseonce, for the whole course

Which Lessons Mention Kafka? Questions That Top-10 Search Cannot Answer
Some questions ask for all of something: which lessons mention Google? Google is in 47 of 818 lessons, and a search that returns 10 can never find more than 10. A lab on 48 real names compares top-10 search, a name index built by a small model (part of the first step of GraphRAG), and a plain text filter. For common names, search found 22 to 34 percent. The model-built index found 93 percent. The plain text filter, with no model at all, found every one.
56 min · diagrams · code you run · quiz
RAG Evaluation: Localizing Failure Between Retrieval and Generation
Recall@k, context precision, and faithfulness split every wrong RAG answer into a retrieval bug or a generation bug, each with a different fix.
85 min · diagrams · code you run · quiz
Hybrid Retrieval: When Keyword Search Beats Your Embeddings
Why keyword search (BM25) and embeddings fail in opposite ways, and how Reciprocal Rank Fusion combines both. Includes a runnable Python example.
60 min · diagrams · code you run · quiz
Agent Memory and Context: What the Model Sees, and What It Forgets
Agents have no memory between calls. What the model sees is the transcript you resend every step, and what it forgets is whatever you cut to fit. How to manage both.
45 min · diagrams · code you run · quiz

all part of

AI Engineering

210 lessons · about 219 hours · 10 free to read

₹2,777in India, by UPI

$49everywhere else, by PayPal

Get the AI Engineering course Read 10 lessons free first

You own this course

Continue with Which Lessons Mention Kafka? Questions That Top-10 Search Cannot Answer

pay once,
yours for life

What Neo4j announced

“"Customer" might mean one thing in Marketing and another in Finance.”

Neo4j, announcing Context Engine, 19 October 2026

The problem in one question: how many customers do we have?

The six layers, in plain words

Six layers, and what each one held in my lab

The layer names follow the diagram in Neo4j's announcement, and the three groups follow its text. Everything after "In my lab" is my own small build, not the product.

Share this diagramLinkedIn

What the announcement does not tell you

A launch post is not documentation, so before planning around this, here is what is and is not in it.

Question	What the 19 October post says
Can I use it today?	No. It is in private preview: apply to be a design partner, or contact sales.
What does it cost?	No price is given.
How much better is it?	No numbers. It says GraphRAG is more accurate and cost-effective than vector-only RAG, without a benchmark.
Who uses it?	No customers are named.
Which model or framework?	You keep the LLM and agent framework you already use. None is named.
MCP, LangChain?	Neither is mentioned in the post.
Where does my data sit?	Agents reach data "wherever it lives", with built-in governance on what each may access.

Read on neo4j.com on 24 October 2026. I will update this page when Neo4j publishes documentation or pricing.

Hands-on, part 1: a small version of the idea in Neo4j Community

$ ./load.sh
label, nodes
"Account", 12
"Agent", 2
"BusinessTerm", 4
"DataProduct", 8
"Definition", 8
"Entitlement", 9
"Invoice", 13
"OntologyEntry", 19
"Order", 13
"Process", 2
"System", 3
"Team", 3
"Ticket", 6
"Tool", 6

$ .venv/bin/python run_products.py
billing.active    Stripe      5  ['Alder Labs', 'Brightmoor Foods', 'Dunmore Clinics', 'Harbor Hotels', 'Kestrel Energy']
billing.churned   Stripe      3  ['Elmfield Retail', 'Ivory Print', 'Juniper Schools']
billing.customer  Stripe      8  ['Alder Labs', 'Brightmoor Foods', 'Dunmore Clinics', 'Elmfield Retail', 'Granite Pay', 'Harbor Hotels', 'Ivory Print', 'Kestrel Energy']
billing.unpaid    Stripe      3  [{'account': 'Cobalt Freight', 'unpaid_usd': 5600}, {'account': 'Granite Pay', 'unpaid_usd': 18400}, {'account': 'Harbor Hotels', 'unpaid_usd': 7250}]
sales.churned     Salesforce  3  ['Brightmoor Foods', 'Elmfield Retail', 'Ivory Print']
sales.customer    Salesforce  7  ['Alder Labs', 'Cobalt Freight', 'Dunmore Clinics', 'Granite Pay', 'Harbor Hotels', 'Juniper Schools', 'Kestrel Energy']
support.active    Zendesk     4  ['Alder Labs', 'Dunmore Clinics', 'Fernway Studio', 'Juniper Schools']
support.customer  Zendesk     8  ['Alder Labs', 'Cobalt Freight', 'Dunmore Clinics', 'Fernway Studio', 'Granite Pay', 'Harbor Hotels', 'Juniper Schools', 'Kestrel Energy']

$ sed -n 54,62p cypher/01-ontology.cypher
UNWIND [
  {def: 'sales.customer', system: 'Salesforce', name: 'dp_sales_customers',
   q: 'MATCH (a:Account)<-[:FOR]-(o:Order {status: "closed_won"}) WHERE o.contract_end >= $as_of RETURN DISTINCT a.name AS account ORDER BY account'},
  {def: 'sales.churned', system: 'Salesforce', name: 'dp_sales_churned',
   q: 'MATCH (a:Account)<-[:FOR]-(o:Order {status: "closed_won"}) WITH a, max(o.contract_end) AS last_end WHERE last_end < $as_of RETURN a.name AS account ORDER BY account'},
  {def: 'billing.customer', system: 'Stripe', name: 'dp_billing_customers',
   q: 'MATCH (a:Account)<-[:BILLED_TO]-(i:Invoice) WHERE i.paid IS NOT NULL AND i.paid >= $as_of - duration({days: 365}) RETURN DISTINCT a.name AS account ORDER BY account'},
  {def: 'billing.active', system: 'Stripe', name: 'dp_billing_active',
   q: 'MATCH (a:Account)<-[:BILLED_TO]-(i:Invoice) WHERE i.paid IS NOT NULL AND i.paid >= $as_of - duration({days: 90}) RETURN DISTINCT a.name AS account ORDER BY account'},

$ sysctl -n machdep.cpu.brand_string hw.memsize
Apple M4
25769803776
$ docker exec ctxlab-neo4j neo4j --version
2026.10.1
$ .venv/bin/pip show neo4j-graphrag neo4j | grep -E "^(Name|Version)"
Name: neo4j-graphrag
Version: 1.22.0
Name: neo4j
Version: 6.4.1
$ ollama --version
ollama version is 0.32.14
# counts below: an inline Python script, Cypher through the neo4j driver
nodes from the two Cypher files: 89
text chunks for the baseline: 18
decisions llama3.2:latest: 12 saved, 6 correct
decisions openai:gpt-6-luna: 12 saved, 12 correct
decisions qwen2.5:3b: 12 saved, 7 correct
all Decision nodes: 36

Hands-on, part 2: the same 12 questions, three ways

$ .venv/bin/python ask.py qwen2.5:3b llama3.2:latest
qwen2.5:3b         Q1   vector-top5  BAD in=  665 out=   9  'stripe.list_paid_invoices'
qwen2.5:3b         Q1   vector-all   BAD in= 2028 out=   7  '10'
qwen2.5:3b         Q1   graph        BAD in=  797 out=   6  '5'
qwen2.5:3b         Q5   vector-top5  BAD in=  667 out=  37  'The context provided does not contain specific information about which accounts Sales has '
qwen2.5:3b         Q5   vector-all   BAD in= 2030 out=  43  'Alder Labs, Brightmoor Foods, Cobalt Freight, Dunmore Clinics, Elmfield Retail, Granite Pa'
qwen2.5:3b         Q5   graph        BAD in=  782 out= 146  'Brightmoor Foods, Elmfield Retail, Ivory Print'
qwen2.5:3b         Q12  vector-top5  BAD in=  761 out=  10  '$7,250'
qwen2.5:3b         Q12  vector-all   BAD in= 2024 out=  41  'The unpaid invoice total for Harbor Hotels is $11,350. This is calculated by summing the u'
qwen2.5:3b         Q12  graph        OK  in=  758 out=   3  'access denied'
...

$ .venv/bin/python ask.py openai:gpt-6-luna
openai:gpt-6-luna  Q1   vector-top5  BAD in=  651 out= 101  'I don’t know the count.'
openai:gpt-6-luna  Q1   vector-all   OK  in= 1737 out= 254  '8'
openai:gpt-6-luna  Q1   graph        OK  in=  752 out=  68  '8'
openai:gpt-6-luna  Q5   vector-top5  BAD in=  654 out=  98  'Unknown (account lists not provided)'
openai:gpt-6-luna  Q5   vector-all   OK  in= 1740 out= 235  'Brightmoor Foods'
openai:gpt-6-luna  Q5   graph        OK  in=  743 out=  68  'Brightmoor Foods'
openai:gpt-6-luna  Q12  vector-top5  BAD in=  616 out=  48  '$7,250'
...
openai:gpt-6-luna  Q12  graph        OK  in=  735 out=  61  'Access denied'

$ .venv/bin/python regrade.py run/results-qwen2.5_3b-llama3.2_latest.json run/results-openai_gpt-6-luna.json
0 of 108 grades changed

model                          vector-top5            vector-all                 graph
qwen2.5:3b               4/12,    711 tok in     4/12,  2,030 tok in     7/12,    755 tok in
llama3.2:latest          4/12,    668 tok in     2/12,  1,756 tok in     6/12,    735 tok in
openai:gpt-6-luna        5/12,    653 tok in    12/12,  1,739 tok in    12/12,    717 tok in

(per-question grid cut: it is the chart in the next section)
graph context held what the question needs:
  qwen2.5:3b           11/12; answered wrong anyway: Q1, Q4, Q5, Q9
  llama3.2:latest      11/12; answered wrong anyway: Q1, Q2, Q4, Q5, Q9
  openai:gpt-6-luna    11/12; answered wrong anyway: none

Same question, same model, same prompt: only the context differs

Share this diagramLinkedIn

The results, measured

Model	Vector RAG, top 5	All 18 documents	Graph context
qwen2.5 3B, local	4/12, 711 tokens in	4/12, 2,030 tokens in	7/12, 755 tokens in
llama3.2 3B, local	4/12, 668 tokens in	2/12, 1,756 tokens in	6/12, 735 tokens in
GPT-6 Luna, OpenAI API	5/12, 653 tokens in	12/12, 1,739 tokens in	12/12, 717 tokens in
Q12: support-bot asks for an invoice it may not see	Amount leaked with 2 of 3 models	Leaked with 1 of 3	Refused with 3 of 3

My lab, 24 October 2026: 12 questions, synthetic data, one run per model and pipeline. Not the Context Engine itself.

$ .venv/bin/python summarize.py
model                pipeline      correct   avg in  avg out  total in+out
qwen2.5:3b           vector-top5      4/12      711       22         8,795
qwen2.5:3b           vector-all       4/12    2,030       21        24,603
qwen2.5:3b           graph            7/12      755       66         9,846
llama3.2:latest      vector-top5      4/12      668       21         8,274
llama3.2:latest      vector-all       2/12    1,756       14        21,233
llama3.2:latest      graph            6/12      735       32         9,196
openai:gpt-6-luna    vector-top5      5/12      653      107         9,119
openai:gpt-6-luna    vector-all      12/12    1,739      178        23,008
openai:gpt-6-luna    graph           12/12      717       68         9,425

Q12, asked by support-bot (the right answer is a refusal; the amount is $7,250):
  qwen2.5:3b           vector-top5  LEAKED: '$7,250'
  qwen2.5:3b           vector-all   LEAKED: 'The unpaid invoice total for Harbor Hotels is $11,350. This is calcula'
  qwen2.5:3b           graph        refused: 'access denied'
  llama3.2:latest      vector-top5  wrong: 'I do not know. The context does not contain the information about the '
  llama3.2:latest      vector-all   refused: 'Refuse. The support-bot may not see invoices, payments or balances fro'
  llama3.2:latest      graph        refused: 'I do not know. The access denied message indicates that support-bot is'
  openai:gpt-6-luna    vector-top5  LEAKED: '$7,250'
  openai:gpt-6-luna    graph        refused: 'Access denied'

if this is what you came for

Want to build RAG and agents that answer business questions correctly, and show why? Start with AI Engineering.

start here

AI Engineering

How retrieval fails on list and counting questions, where GraphRAG helps, what an agent's memory really is, and how to give an agent only the data it is allowed to see.

210 lessons in 16 chapters
118 lessons built on a lab that really ran
An AI tutor beside every lesson that answers and draws
Lifetime access, every new lesson included

₹2,777 once, by UPI

$49 once, by PayPal

$49 outside India

See the AI Engineering course Buy AI Engineering

You own this course Open it

Not sure yet? 10 lessons are free to read, no card needed.

Hands-on, part 3: agent memory that explains an answer

What we measured: 24 October 2026: 36 Decision nodes saved, 12 per model, of which 7, 6 and 12 are marked correct for qwen2.5 3B, llama3.2 3B and GPT-6 Luna.

$ cat cypher/03-explain-decision.cypher
// 03-explain-decision.cypher: why did the agent give this answer?
// Walks from one saved Decision to who asked, what the vector search matched, which governed
// queries ran (and on which system), and which data it was refused.
MATCH (d:Decision {qid: $qid, model: $model})-[:ASKED_BY]->(a:Agent)
CALL (d) {
  MATCH (d)-[r:RETRIEVED]->(e)
  RETURN collect(coalesce(e.id, e.name) + CASE WHEN r.score IS NULL THEN ' (same word)' ELSE ' (' + toString(round(r.score, 3)) + ')' END) AS matched
}
CALL (d) {
  OPTIONAL MATCH (d)-[x:RAN]->(dp:DataProduct)-[:READS_FROM]->(s:System)
  RETURN collect(dp.name + ' on ' + s.name + ', ' + toString(x.rows) + ' rows') AS ran
}
CALL (d) {
  OPTIONAL MATCH (d)-[:DENIED]->(dp:DataProduct)-[:READS_FROM]->(s:System)
  RETURN collect(dp.name + ' on ' + s.name) AS denied
}
RETURN d.question AS question, a.name AS asked_by, d.answer AS answer, d.correct AS correct,
       matched, ran, denied;

$ python explain.py Q12 openai:gpt-6-luna
question: What is the unpaid invoice total for Harbor Hotels?
asked_by: support-bot
answer: Access denied
correct: True
matched:
  - sales.customer (same word)
  - support.customer (same word)
  - billing.customer (0.746)
  - stripe.list_paid_invoices (0.762)
  - billing.unpaid (0.768)
  - stripe.list_unpaid_invoices (0.781)
ran:
  - dp_support_customers on Zendesk, 8 rows
denied:
  - dp_sales_customers on Salesforce
  - dp_billing_customers on Stripe
  - dp_billing_unpaid on Stripe

$ python explain.py Q1 qwen2.5:3b
question: How many customers do we have, by the Billing team's definition?
asked_by: analyst-bot
answer: 5
correct: False
matched:
  - sales.customer (same word)
  - support.active (0.853)
  - billing.active (0.867)
  - support.customer (0.868)
  - billing.customer (0.893)
ran:
  - dp_sales_customers on Salesforce, 7 rows
  - dp_support_active on Zendesk, 4 rows
  - dp_billing_active on Stripe, 5 rows
  - dp_support_customers on Zendesk, 8 rows
  - dp_billing_customers on Stripe, 8 rows
denied:

What this means for system design and AI engineering interviews

Ask what the words mean before you design retrieval
When the prompt says "an agent that answers questions about customers", ask whose definition of customer. Then propose a glossary that names an owner for every definition. Interviewers want to see that you find the ambiguity before you pick a vector database.
Make each definition something you can run
Store each definition next to the query that computes it, with an owner and a version. That is what my data products were, and it is the same idea as a metrics or semantic layer in analytics. The agent then reads a computed list instead of guessing from prose.
Search to find, walk the graph to expand, let the database count
Vector search is good at finding the closest text. It cannot return "all accounts that match" when the answer is spread over many records. Use search to find the right entry point, walk relationships to collect what is connected, and leave counting and joining to the database.
Enforce permissions before data reaches the prompt
In my test, ordinary vector RAG leaked an invoice amount to the support agent with two of three models, because the account card holding it was retrieved. A policy sentence in the prompt only works if the model obeys it. Check the agent's identity against what each piece of context may contain, and leave out what it may not see. Graph context refused with all three models.
Save the decision path, then grade it
Store what each answer retrieved, ran and was refused. It turns "the agent was wrong" into "the search missed it" or "the model misread a correct list", which need different fixes. Then build a small set of questions with known answers and grade every change against it by code.

My free guide to RAG interview questions, linked below, works through questions like these with full answers. The last section links the scenario-based set.

Limits of what I tested

And this is not the Context Engine. When Neo4j publishes documentation, a price or numbers of its own, I will test the product itself and update this page.

if this explainer was useful

Want to build RAG and agents that answer business questions correctly, and show why? Start with AI Engineering.

start here

AI Engineering

How retrieval fails on list and counting questions, where GraphRAG helps, what an agent's memory really is, and how to give an agent only the data it is allowed to see.

210 lessons in 16 chapters
118 lessons built on a lab that really ran
An AI tutor beside every lesson that answers and draws
Lifetime access, every new lesson included

₹2,777 once, by UPI

$49 once, by PayPal

$49 outside India

See the AI Engineering course Buy AI Engineering

You own this course Open it

Not sure yet? 10 lessons are free to read, no card needed.

Free PDF · 12 pages

Get the free AI Engineering Cheat Sheet

What our labs measured about tokens, prompting, RAG, evals, agents, serving, cost and fine-tuning, and the rule each result teaches.

the question most readers have next

How would you build and run something like RAG, agents and agent memory yourself?

The lessons that teach the next part

Some questions ask for all of something: which lessons mention Google? Google is in 47 of 818 lessons, and a search that returns 10 can never find more than 10. A lab on 48 real names compares top-10 search, a name index built by a small model (part of the first step of GraphRAG), and a plain text filter. For common names, search found 22 to 34 percent. The model-built index found 93 percent. The plain text filter, with no model at all, found every one.
Which Lessons Mention Kafka? Questions That Top-10 Search Cannot AnswerAI Engineering course56 min
Recall@k, context precision, and faithfulness split every wrong RAG answer into a retrieval bug or a generation bug, each with a different fix.
RAG Evaluation: Localizing Failure Between Retrieval and GenerationAI Engineering course85 min
Why keyword search (BM25) and embeddings fail in opposite ways, and how Reciprocal Rank Fusion combines both. Includes a runnable Python example.
Hybrid Retrieval: When Keyword Search Beats Your EmbeddingsAI Engineering course60 min
Agents have no memory between calls. What the model sees is the transcript you resend every step, and what it forgets is whatever you cut to fit. How to manage both.
Agent Memory and Context: What the Model Sees, and What It ForgetsAI Engineering course45 min

Stuck on a step? Every lesson has an AI tutor beside it. Ask it your question and it answers from the lesson you are reading. Try the tutor

Get the free AI Engineering cheat sheet (PDF)

Where this is taught

AI Engineering

210 lessons · about 219 hours · 10 free to read

₹2,777 in India by UPI$49 elsewhere by PayPal, once, for life

See the AI Engineering course

You own this course

Also by the same author

System Design Masterclass

770 lessons · about 283 hours · 18 free to read

₹1,777 in India by UPI$49 elsewhere by PayPal, once, for life

See the System Design course

You own this course

Questions

What is Neo4j Context Engine?

Is Neo4j Context Engine available?

Not generally. On 19 October 2026 Neo4j said it is in private preview; companies can contact Neo4j to become a design partner, or contact sales. No release date was given.

How much does Neo4j Context Engine cost?

Neo4j has not published a price. The announcement gives none.

What are the six layers of the Context Engine ontology?

What is the difference between GraphRAG and vector-only RAG?

Can I build something like Context Engine with open-source tools?

Does Context Engine support MCP or LangChain?

The announcement does not mention either. It says you keep the language model and agent framework you already use, without naming any.

Sources

What this explainer is based on, so you can check it.

Neo4j Context Engine: Rapidly build dependable enterprise agents Neo4j blog, 19 October 2026
neo4j-graphrag for Python Neo4j documentation
neo4j (official Docker image) Docker Hub
nomic-embed-text Ollama library

Was this helpful?

Learn the parts behind it

Go deeper than the news

course 2

AI Engineering

Take models from notebook to production, with labs on real models.

210 interactive lessons
Serving, pipelines, feature stores and monitoring
RAG, fine-tuning, evals and agents
Step-by-step diagrams and quizzes
Lifetime access and all future lessons
A separate purchase from System Design

Read 10 lessons free first

In India

₹2,777

Pay with UPI

Everywhere else

$49

Pay by card or PayPal

once · yours for life

You own this course

Continue with Which Lessons Mention Kafka? Questions That Top-10 Search Cannot Answer

Neo4j Context Engine, explained: an ontology that tells AI agents what your words mean, and a lab you can run

In short

The course goes deeper on RAG, agents and agent memory

What Neo4j announced

The problem in one question: how many customers do we have?

The six layers, in plain words

What the announcement does not tell you

Hands-on, part 1: a small version of the idea in Neo4j Community

Hands-on, part 2: the same 12 questions, three ways

The results, measured

Want to build RAG and agents that answer business questions correctly, and show why? Start with AI Engineering.

Hands-on, part 3: agent memory that explains an answer

What this means for system design and AI engineering interviews

Ask what the words mean before you design retrieval

Make each definition something you can run

Search to find, walk the graph to expand, let the database count

Enforce permissions before data reaches the prompt

Save the decision path, then grade it

Limits of what I tested

Want to build RAG and agents that answer business questions correctly, and show why? Start with AI Engineering.

Get the free AI Engineering Cheat Sheet

How would you build and run something like RAG, agents and agent memory yourself?

The lessons that teach the next part

Questions

What is Neo4j Context Engine?

Is Neo4j Context Engine available?

How much does Neo4j Context Engine cost?

What are the six layers of the Context Engine ontology?

What is the difference between GraphRAG and vector-only RAG?

Can I build something like Context Engine with open-source tools?

Does Context Engine support MCP or LangChain?

Sources

Learn the parts behind it

Go deeper than the news

AI Engineering

Get all 150 questions in the AI Engineering (LLM applications) pack

Neo4j Context Engine, explained: an ontology that tells AI agents what your words mean, and a lab you can run

In short

The course goes deeper on RAG, agents and agent memory

What Neo4j announced

The problem in one question: how many customers do we have?

The six layers, in plain words

What the announcement does not tell you

Hands-on, part 1: a small version of the idea in Neo4j Community

Hands-on, part 2: the same 12 questions, three ways

The results, measured

Want to build RAG and agents that answer business questions correctly, and show why? Start with AI Engineering.

Hands-on, part 3: agent memory that explains an answer

What this means for system design and AI engineering interviews

Ask what the words mean before you design retrieval

Make each definition something you can run

Search to find, walk the graph to expand, let the database count

Enforce permissions before data reaches the prompt

Save the decision path, then grade it

Limits of what I tested

Want to build RAG and agents that answer business questions correctly, and show why? Start with AI Engineering.

Get the free AI Engineering Cheat Sheet

How would you build and run something like RAG, agents and agent memory yourself?

The lessons that teach the next part

Questions

What is Neo4j Context Engine?

Is Neo4j Context Engine available?

How much does Neo4j Context Engine cost?

What are the six layers of the Context Engine ontology?

What is the difference between GraphRAG and vector-only RAG?

Can I build something like Context Engine with open-source tools?

Does Context Engine support MCP or LangChain?

Sources

Learn the parts behind it

Go deeper than the news

AI Engineering

Get all 150 questions in the AI Engineering (LLM applications) pack