Agentic RAG
The knowledge base the agent reads before it answers.
This service needs the key in its OWN env, and its failure is silent: ingest answers 200 and embeds nothing. Save the key once in OpenAI settings
documents in the base: 0
The base is empty — add a document to have something to ask about.
The knowledge graph — relations, compared with vector search
What you get. A composed answer, not a list of passages. The service reads across documents and writes the reply itself.
Why that is. Two phases. At ingest a model extracts entities and the relations between them into a graph; at question time that graph is walked. The thinking is paid for in advance, once per document, instead of at every question.
Where it wins. When the answer is spread across many documents and depends on how facts connect: who is related to whom, what follows from what, summaries over a whole corpus.
What it costs. Noticeably more than vectors, and mostly at ingest: every document is read by a model. A question is cheap by comparison, but never free — it also goes through a model.
Where it is weak. Changing data. A corrected document has to be re-read from scratch, relations and all, while a vector store needs one embedding call. And the graph decides itself what mattered in your text — you cannot see the passage its answer came from as directly.
The two do not share storage. The graph lives in this service's own folder; vectors live in the data service's SQLite. To have a document in both, add it to both: you pay each one's ingest and get each one's kind of answer. Nothing is shared but the OpenAI key.