Weave
Knowledge portal and chat with sources
Important answers often sit in documents nobody can find, or with people who are not available. Weave reads PDFs, Office files, scans, email and Confluence pages, turns them into structured text and answers questions in a chat. Every answer includes its source, so people can check where a statement came from.
Three rules keep knowledge findable without giving up control: nothing is indexed without an explicit release. Permissions take effect inside the search query, not afterwards. And without sources, there is no answer. When the search finds nothing, the chat says so. Everything runs on your own hardware, the language model is interchangeable, and no external service is connected out of the box.
How it works
Agent mode: multiple subagents research in parallel
For complex questions, a planner splits the question into sub-questions and starts a configured subagent for each, every one with its own knowledge space and search budget. The chat shows progress before the merged answer with sources appears.
Features
- Knowledge spaces group documents; the groups allowed to read them are set there
- Release as an immutable snapshot by the owner or an administrator; only then is anything indexed
- Hybrid search: vector search over pgvector and PostgreSQL full text, merged via RRF and reordered by a cross-encoder
- Permissions through groups, evaluated inside the search query — outside a knowledge space there are no results, not even hidden ones
- Chat with a source reference for every answer, using the same sign-in as the portal, local or via OIDC
- Agent mode: for complex questions, several configured subagents research in parallel within their knowledge spaces, with a progress indicator in chat
- Bots via n8n with streaming for long-running research; MCP server and REST mirror for list_collections and search with short-lived scopes for flows, plus an importable n8n sample flow for Weave RAG with an external OpenAI-compatible LLM
- Technical identities for external MCP/REST integrations: explicitly granted knowledge spaces, a one-time token, rotation, revocation and an audit log
- SSO sign-in for MCP clients such as VS Code (Entra ID, Keycloak): the access token maps to an existing user, whose permissions come from their groups in Weave
- Managed bots and group knowledge spaces; permission and content changes synchronize through search and chat
- Released document versions can be withdrawn and are then removed from the search index; "reindex" is available per document and per knowledge space
- Processing profiles for OCR (PP-OCRv6, PP-StructureV3) and configured vision models, with reprocessing after the quality check
- Quality grades with an explanation in the review portal, filterable; provenance per document and who released it are visible; bulk actions for multiple documents
- Confluence import with refresh and images shown directly in chat answers, plus single .eml files as sources, with supported attachments processed alongside
- Disaster recovery: encrypted export of all Ingest data, restore with automatic reindexing; admin index maintenance (recompute vectors, rebuild from releases)
- Self-hosted via Docker Compose or as the complete platform through the published Helm chart; optional embedding and reranking services run on CPU
- Horizontally scalable: every service and both worker pools (OCR, import/release) run as several replicas, files live in PostgreSQL instead of on a shared volume, and interrupted jobs are requeued automatically
Write to me
Whether you have product feedback or want to discuss working together, your message comes straight to me. I reply myself.