Ask a question about technical documentation and the answer often sits in three places at once: a release note, a deprecated feature page and the doc that replaced it. At this month’s Agentic AI Context Builders meetup, Pratham Batra, AI team lead at Arango, showed how Arango AutoGraph Studio turns a pile of docs into a context graph that answers those questions with citations, and keeps it current as the docs change.
The session also introduced Nathan Loding, Arango’s new developer advocate, who takes over the meetup from November.
Watch the full session:
In this recap: why docs are a graph problem, three live demos, and the best questions from the audience.
Technical documentation is already a graph
Pratham opened with a simple point: docs link to each other. A release note deprecates a feature, the feature page explains it, and a later release note removes it. Kubernetes does the same when PodSecurityPolicy gives way to Pod Security Admission.
So a question like “Can I still run Pregel on ArangoDB 3.12?” needs three pages at once. Vector search ranks each chunk on its own. A graph follows the links between pages.
Demo 1: from a folder of docs to a context graph
In AutoGraph Studio, Pratham created a project and uploaded folders of ArangoDB documentation: connectors, security, the Kubernetes operator, the storage engine and the query engine. Each folder became a category that helps guide how the graph is built. He then chose his models. Any OpenAI-compatible endpoint works, including OpenRouter or a model you host yourself.
AutoGraph builds the context graph in two layers:
- Corpus graph: how documents relate to each other, grouped into domains and clusters.
- Knowledge graph: the entities and relationships inside those documents, grouped into communities.
Between the two steps you choose how much of the corpus gets full GraphRAG and how much stays VectorRAG, with a cost estimate on screen. Pratham’s advice: “Everything deserves a graph,” but the slider lets you spend where relationships matter. Every entity keeps a link back to the document it came from.
Demo 2: vector RAG vs. AutoGraph on a multi-hop question
Pratham asked the same question of a basic vector RAG app (cosine similarity over chunks) and of AutoGraph: why can’t a bank’s compliance team see which service wrote each record, and how do they fix it?
Vector RAG found the problem but said the context “does not state the fix.” The fix lives in a document that never mentions the bank, so similarity search can’t reach it. AutoGraph followed the relationships and returned the fix, step by step, with a citation on every claim.
Two search modes cover different questions:
- Instant Search answers direct questions quickly and at lower cost.
- Deep Search breaks a hard question into sub-queries, picks a retriever for each and shows its reasoning in a Trace tab.
Click any citation to see the exact chunk, open it in the database, or view the subgraph used for the answer.
Demo 3: keeping the graph current without a rebuild
Docs change, so the graph has to change with them. Pratham added three new documents, including the release notes for a new version of the sample product, through incremental graph updates. Only the new documents were processed, and the graph reorganized its communities around them.
| Before | After | |
|---|---|---|
| Entities | 1,482 | 1,639 |
| Communities | 159 | 170 |
The retriever then answered questions that only the new release notes could answer, such as when a removed feature stopped working. You can also delete documents the same way. Incremental updates run through the API today; a UI is on the way.
From the Q&A
- Corpus graph or knowledge graph? The corpus graph links documents; the knowledge graph links the entities inside them. Together they form the context graph.
- Why two build steps? So you can set the GraphRAG strategy, and the cost, before the knowledge graph is built.
- Do entities link back to their source? Yes. Every entity carries metadata pointing to its document and cluster, which powers the citations.
- Is there an API? Yes, alongside the UI, plus a Python SDK.
- Does one query cover vector and graph? Yes. AQL, the ArangoDB query language, runs vector, BM25 keyword and graph search in a single query. See the AQL docs.
- Benchmarks against out-of-the-box RAG and GraphRAG? The team will share results in a follow-up.
One attendee summed up where graphs shine: a ticketing system where issues, types and owners are all connected. Connecting a new issue to one solved years ago is exactly what a context graph does well.
Try Arango AutoGraph Studio yourself
Everything in the demos is in the free Developer Edition of the Arango Contextual Data Platform: ArangoDB, the Platform Suite and the Agentic AI Suite, including AutoGraph Studio. A managed version is coming soon.
- Developer Edition: arango.ai/arango-developer-edition
- Arango AI SDK for Python: github.com/arangoml/arango-ai-sdk
- Discord community: arango.ai/join-discord, where Nathan and the team answer questions
- Step-by-step tutorial: “Build with AutoGraph” at docs.arango.ai
Thanks to Pratham and everyone who joined and asked questions. Have a topic you want covered next? Tell us on Discord or the Agentic AI Context Builders meetup page.
