Docs Knowledge Graph
GraphRAG over technical documentation: semantics plus structure
A retrieval layer that turns a large technical documentation set into a property graph alongside a vector index, so retrieval can follow structure (prerequisites, versions, cross-references) and not just semantic similarity.
- Role
- Sole developer: pipeline, graph schema, retrieval design
The problem
Plain vector search over documentation retrieves passages that sound like the question. That fails on exactly the questions that matter in technical docs: what must I understand first, which release does this apply to, and where else is this concept defined. Similarity has no notion of prerequisite or version.
The approach
01
Extract structure deterministically
Documentation is already structured: headings, release families, cross-links, API references. Rather than asking a model to infer that structure and hoping it is consistent, the pipeline extracts it deterministically into a typed property graph.
02
Run both indexes, use both
A vector index handles 'what is relevant to this question'. The graph handles 'what is connected to it, and in which version'. Hybrid retrieval combines the two rather than picking a side.
03
Model multiple sources in one schema
Prose documentation and a typed SDK reference describe the same platform in different shapes: concepts and tasks on one side, constructors and tables on the other. A multi-source schema keeps them linked instead of siloed.
How it turned out
- In production as Morgan's retrieval layer: every grounded answer at studywithmorgan.com is served through this graph.
- Indexes 514,750 documentation chunks across multiple release families, plus a typed SDK reference, in a single queryable graph.
- Fully self-contained: owns its own storage, embedding, and retrieval interface, and stays deployable independently of any consuming application.
Want something like this built?