kglite-docs

Agent-first knowledge base for documents. Ingest PDFs, Office files, Markdown, HTML, or images; chunk + embed them with BAAI/bge-m3; cluster, tag, summarise, fact-check, translate, and review them — and serve the whole thing to AI agents over MCP.

Built on kglite (storage + vector search + clustering) and mcp-methods (MCP framework).

Why this and not generic RAG?

Most "RAG libraries" hand the agent search(query) → list[chunk] and stop. kglite-docs treats the corpus as a living knowledge graph that records who did what — and gives the agent typed tools to act on it.

• 📄 Multi-format ingest — PDF, DOCX, PPTX, MD, HTML, TXT, images. All flow into the same Document → Page → Chunk shape.
• 🤝 Agents are first-class nodes — their views, tags, summaries, verifications, and reviews are all queryable.
• ✅ Cross-checked summaries — one agent writes, a different agent verifies. Self-verification is rejected server-side.
• 📋 Review kanban — chunks move through new → in_review → reviewed with an immutable audit trail.
• 🛡️ Grounding checks — score how well an agent's summary aligns with its sources. Catch hallucinations before they ship.
• 🌍 Translations — per-chunk, multi-translator, with author/reviewer provenance.
• 🖼️ Agent-driven OCR — scanned pages handed back as rendered PNGs; agent transcribes and the graph absorbs the result.
• 🔒 Local & private — parsing, embedding, and analysis all run on your machine against a local .kgl file. The only network call is a one-time bge-m3 model download; your documents never leave the host. See Confidentiality.

Install

pip install kglite-docs

30 seconds of Python

from kglite_docs import Corpus

with Corpus.create("kb.kgl") as corpus:           # auto-saves on exit
    corpus.ingest_dir("./papers")                  # PDF / DOCX / PPTX / MD / HTML / images
    hits = corpus.search("transformer attention", top_k=5, agent_id="me")
    ctx = corpus.compose_context("transformer attention", max_tokens=3000)
    # ctx["items"] is a ranked, token-budgeted bundle ready for your LLM prompt

30 seconds of agent loop

Cross-checked enrichment in five lines:

sid = corpus.add_summary(
    target_id=hits[0]["id"], text="DPR uses a dual BERT encoder…",
    agent_id="writer", model="opus-4.7",
)
# A different agent verifies — self-verification is rejected
corpus.verify_summary(sid, verdict="verified",
                      verifier_agent_id="reviewer", notes="checked p.5")
# Score how grounded the summary is in its source chunks
print(corpus.check_grounding(sid)["supported_fraction"])    # → 1.0

Run it as an MCP server

kglite-docs-mcp --db kb.kgl

Register with Claude Code:

claude mcp add kglite-docs -- kglite-docs-mcp --db /abs/path/kb.kgl

The agent now sees ~30 typed tools (search, compose_context, add_summary, verify_summary, tag_chunk, cluster_chunks, claim_next_review, …) plus cypher_query as an escape hatch.

Read the docs

📖 Full documentation at kglite-docs.readthedocs.io

• Getting started — 10 minutes from pip install to a running agent
• Agent workflows — research, comparison, fact-checking, OCR loops, hallucination guards
• Architecture — graph model, design rationale, the 30+ typed MCP tools
• API reference — every method, every argument, IDE-friendly type stubs
• Troubleshooting — common failure modes
• Confidentiality — everything runs local; what the one network call is (and isn't)
• Changelog

License

MIT.