Open source · Vectorless retrieval engine
Logidoc is an open-source, self-hosted document indexing server for AI agents. No vectors, no embeddings. Upload a PDF, get a searchable table of contents. Plug-and-play via MCP for any agent client, with SDKs in Go, TypeScript, and Python.


The pain
Vector-based RAG pipelines are expensive, opaque, and their relevance depends on arbitrary chunking. PageIndex points in a more reasonable direction but it's heavy to run and pricey when hosted. There's room for something lighter that a dev can spin up locally to prototype an AI agent against a stack of long documents.
The approach
Reasoning-based retrieval. The agent navigates a hierarchical table of contents and pulls the exact section text. No embeddings, no chunking, full traceability. Cost stays controllable, install is trivial, integration is plug-and-play.
What's hard
The real challenge sits in PDF parsing, specifically tables and embedded images. Quality varies a lot with the LLM model used, and OCR fallbacks like Tesseract are not great without a trained recognition model on top. This is the part that still needs work.
Status
In development. The MCP server works, the three SDKs ship, the docs site is live. What remains is performance work, benchmarks, full observability, a real debug interface.
Roadmap
LLM tracing via Phoenix. Prometheus and Grafana dashboards. LLM cost monitoring. A proper prompt-evaluation client. Auto-generated Python and TypeScript SDKs.