Beever Atlas

Beever Atlas turns the conversations a team already has on Slack, Discord, Microsoft Teams, and Mattermost into a knowledge base that maintains itself. A six-stage pipeline fetches messages in real time, extracts atomic facts, entities, and relationships with a model, filters extractions below a confidence threshold, writes them to storage through an outbox pattern, clusters related facts by cosine similarity, and generates wiki pages with citations and diagrams. PDFs, Markdown, and other documents can be ingested alongside chat messages.

Two memory systems back retrieval. A three-tier semantic store — channel, topic, and atomic fact — handles meaning-based lookup in Weaviate, and a graph store in Neo4j maps entities and their relationships for multi-hop and temporal questions such as who decided something or how an approach evolved. A query router picks a strategy per question.

Retrieval runs against the distilled wiki rather than raw chat history. The expensive distillation happens once, at ingestion, so a query hits deduplicated, topic-organized prose instead of fragmented turn-by-turn messages, and every claim links back to the source messages that produced it. The QA agent falls back to raw messages only when a fact has not been distilled yet. The wiki is also usable on its own: a channel page carries an overview, concept map, topics, people, decisions, FAQ, and glossary that someone joining a channel can read instead of scrolling months of history. The approach follows the observation that models reason better over encyclopedic content than over conversational transcripts.

Answers stream to a dashboard over server-sent events, or reach external coding agents through an MCP server at /mcp exposing 28 tools for discovery, retrieval, wiki reading, graph traversal, and long-running operations, with per-agent API keys and principal-keyed rate limits. A standalone stdio mode exposes the same tool catalog with no HTTP server or backing stores.

The stack runs as Docker Compose services — a FastAPI backend with ADK agents, a platform bridge bot, and a React dashboard — over Weaviate, Neo4j, MongoDB, and Redis. Embedding and agent providers are chosen at setup, and platform credentials are entered through the web UI and stored encrypted rather than in environment files. Media shared in synced channels is persisted so it keeps rendering after the platform CDN link expires. No telemetry is collected. The /api/* endpoints are marked unstable in 0.1.0, with a versioned prefix planned for 0.2.0.