Get Started
Overview
Turn any codebase into a graph-aware vector database that local AI agents can read, search and document — using a fraction of the tokens.
The name
dokkai (読解) is Japanese for “reading comprehension” — which is exactly what this gives an LLM over your code.What is Dokkai?
Dokkai ingests a repository, builds a full dependency graph of it (calls, inheritance, definitions, modules), slices the real source into rich chunks, and stores everything in a vector database. On top of that it serves a graph-augmented RAG: instead of returning a handful of loosely-matched snippets, it retrieves the connected neighbourhood of the code you asked about, so a local LLM can answer accurately and document parts of the system that were never documented.
Everything runs 100% locally — Weaviate for vectors, Ollama for both embeddings and generation. Your code never leaves your machine.
Ask a question, get a grounded answer streamed back
bash
curl -N -X POST localhost:8000/chat \
-H "Authorization: Bearer $TOKEN" \
-H 'Content-Type: application/json' \
-d '{"message":"how does a request move through the system?","project_name":"your-repo","audience":"developer"}'Principles
Graph-augmented
Retrieval walks the dependency graph — callers, callees, definitions — not just the nearest vectors.
Token-lean
Agents pull the exact connected context they need instead of paying to read whole files.
Local & private
Weaviate + Ollama run on your machine. Nothing is sent to a third party.
Idempotent
Deterministic UUIDs mean re-ingestion upserts entities in place — no duplicates.
Why it matters
- Token economy. Agents grep and read whole files to understand code. Dokkai lets them pull the exact connected context semantically.
- Automatic documentation. Generate complete, audience-specific docs (developer / manager / customer) for systems that were never documented.
- Semantic search for agents. Ask “how does a request move through the system?” and get the relevant subsystem — not keyword hits.