Introduction

A self-hosted knowledge platform built on a typed knowledge graph, fused retrieval, semantic memory, and first-class database connections.

Fabric is a self-hosted knowledge platform. Connect Gmail, Slack, Drive, Fireflies, IMAP, and your databases. Fabric builds a typed knowledge graph across everything and lets you chat with the whole stack through one interface.

Typed knowledge graph

Ten typed edge kinds. Not flat document search — real relational reasoning.

Fused retrieval (RRF)

BM25 + vector, combined by rank position. Same pattern as Vespa and Elastic.

Memory with decay

Strengthens what you use. Prunes what you don’t. Grounded in the graph.

Databases as first-class

Real SQL against Postgres and MySQL. Not API wrappers.

Claude-native

Agent SDK, streaming responses, tool use, session persistence.

Langfuse tracing

Every LLM call, tool, embedding, and cost per query.

Ask across your entire stack

The questions most AI assistants can’t answer — because they’re built on document search, not relational knowledge.

Who has context?

Q: Who has context on the Q2 billing migration?

Fabric traverses its graph across email threads, Slack #billing, and the last three Fireflies meetings. Returns four people with evidence — not documents.

Join SQL + Slack

Q: Pull customers from Postgres public.customers who opened a support ticket in the last 7 days, and match them to any Slack #support threads that mention them.

Real SQL runs against your real database; results land in chat next to the relevant Slack threads. Dust can’t do this. Glean can’t either.

Date-range timeline

Q: What happened with Acme between January and March?

Every node and edge carries a source_date. Fabric filters the graph by date range across email, meetings, Slack, and docs, then returns a timeline.

Four layers that compound

1 — Typed knowledge graph

Nodes are people, threads, meetings, channels, customers, domains. Edges are typed — sent_by, replied_to, attended, participant, in_thread, posted_in, organized_by, in_folder, from_domain, has_email — each carrying a weight and a timestamp.

Graph traversal is a first-class query operation, not a post-hoc extraction step.

2 — Fused retrieval via RRF

Every query runs BM25 (Postgres ts_rank_cd) and vector similarity (pgvector cosine) in parallel, then combines them by rank position via Reciprocal Rank Fusion with k = 60.

rrf_score = 1 / (60 + vec_rank) + 1 / (60 + bm25_rank)

Most RAG tools still do weighted linear blending on incompatible score scales, which quietly lets one signal dominate. We don’t.

3 — Semantic memory with decay

Every conversation produces typed observations — facts, decisions, commitments, risks, insights, patterns. Each has an importance score that strengthens on reference (×1.1) and decays when unused (×0.9 per conversation). Below 0.05, pruned.

Unlike mem0’s floating memory, every observation links back to the email, meeting, or message where the fact came from.

4 — Databases as first-class citizens

Direct connections to PostgreSQL and MySQL via asyncpg and aiomysql with full schema discovery. Natural-language questions become real SQL against your real database.

Most “chat with your data” tools wrap SaaS APIs. Fabric connects to the actual systems where your operational data lives.

Who Fabric is for

Technical teams that want to own their stack. Run the backend yourself. Pick your model. Inspect the graph in SQL. Read the source. Modify what you need.

Self-host anywhere

Docker Compose on a laptop or ECS Fargate via AWS Copilot. Both first-class.

Pick your model

Anthropic, OpenAI, or local — switchable per deployment.

SQL-inspectable

Graph and memory live in Postgres. No black box.

Traced end-to-end

Langfuse shows every LLM call, tool use, and per-query cost.

Get started

Quickstart

Docker Compose to a running Fabric in 10 minutes.

Architecture

The end-to-end data flow.

How It Works

The engineering behind each of the four layers.

API Reference

72+ endpoints, live OpenAPI playground.