beever-atlas

Beever Atlas pulls the conversations your team already has on Slack, Discord, Microsoft Teams, and Mattermost, extracts atomic facts, deduplicates them, and clusters them into topic pages with citations. A graph store links the people, decisions, and projects mentioned across channels. Ask questions in natural language and get answers cited back to the source messages — through the dashboard, or through MCP into Claude Code and Cursor.

If you want a knowledge base that grows on its own from the chats your team already has, this is it.

✨ Features in action

Six short clips — connect a workspace, sync history, watch memory build, browse the auto-generated wiki, ask questions, plug external AI agents in via MCP.

🏗️ Architecture

Conversations from any supported platform flow into a unified ingestion pipeline that produces two complementary memory systems — a 3-tier semantic store (channel / topic / atomic fact) for fast hybrid search, and a graph store that extracts entities and their relationships. Those memories fuel two consumer surfaces: the LLM Wiki (distilled, auto-maintained) and QA Agents (served through the dashboard directly, or through MCP into Claude Code / Cursor).

Under the hood, three services (backend, bot, frontend) are backed by four data stores (Weaviate, Neo4j, MongoDB, Redis). See the architecture overview on the documentation site for the full design — component responsibilities, dual-memory internals, and the smart query router.

💡 Why Wiki-First RAG?

Most RAG systems answer questions by retrieving raw message snippets and feeding them straight to an LLM. Beever Atlas takes a different approach: it continuously distils conversations into a structured, auto-maintained wiki — with topic pages, entity graphs, decisions, and citations — before any query is issued. When you ask a question, the retrieval layer works against clean, deduplicated knowledge rather than noisy chat history. This means answers are more consistent, citations are traceable to source messages, and the wiki itself becomes a useful artifact your team can browse independently of the Q&A interface. The dual-memory architecture (semantic + graph) lets the query router pick the right retrieval strategy per question, keeping latency low and context precise.

The inspiration: LLMs read wikis, not chat logs

The per-channel wiki concept is directly inspired by Andrej Karpathy's observation that LLMs are far better at reasoning over curated, encyclopedic content (books, docs, wikis) than over raw conversational transcripts. Chat history is noisy, redundant, temporally scattered, and full of implicit context that only humans resolve. A wiki, by contrast, is the already-distilled form of that knowledge — deduplicated, structured, citation-bearing, and organised by topic rather than by timestamp.

Beever Atlas operationalises this insight: every synced channel gets its own auto-generated, continuously-updated wiki — sections for topics, entities, decisions, open questions, and timelines — rebuilt incrementally as new messages arrive. The QA agent retrieves against this wiki first, falling back to raw messages only when a fact hasn't been distilled yet.

What this unlocks in practice

• Better answers, fewer hallucinations — retrieval operates on fact-dense prose with explicit entity relationships, not on fragmented turn-by-turn chat.
• Traceable citations — every wiki claim links back to the source messages that produced it, so answers are auditable all the way down to the original Slack/Discord/Teams thread.
• A browsable artifact, not just a Q&A box — the wiki is useful on its own. New teammates onboarding to a channel can read the distilled wiki instead of scrolling three months of history.
• Cheaper inference at query time — the expensive distillation work happens once, at ingestion. Queries hit compact, pre-digested context instead of re-summarising raw logs on every request.
• Graph-aware reasoning — the entity graph built alongside the wiki lets the query router answer relational questions ("who worked on X with Y?") that pure vector RAG struggles with.

For a detailed comparison with other LLM knowledge tools, see the comparison page on the documentation site.

🚀 Quick Start

Beever Atlas ships as a Docker Compose stack (backend + bot + web + 4 datastores). You can try a seeded demo in 30 seconds with zero keys, then pick one of three deployment options to install it for real.

1. Get the code

git clone https://github.com/beever-ai/beever-atlas.git
cd beever-atlas

2. Try the demo first (optional, no keys needed for seeding)

make demo

make demo brings up the full stack pre-loaded with a public Wikipedia corpus (Ada Lovelace + Python history). Seeding uses pre-computed