Knowledge Fabric

Knowledge Fabric is a vendor-neutral evidence retrieval platform.

In plain terms: it helps an assistant find the right context before generating a response. Instead of jumping straight to “an answer,” it returns a structured evidence package that downstream assistants can use for grounded output.

Why this project exists

Many assistant failures come from missing or weak context, not from poor generation quality.
Knowledge Fabric focuses on that retrieval layer so teams can improve relevance, traceability, and trust.

Related MCP server: MCP Knowledge Service

What it does (Phase 1)

• Ingests source content (markdown, text, html, pdf, docx, pptx)

• Chunks and normalizes content for retrieval

• Supports lexical search, vector search, and hybrid fusion

• Returns evidence packages through MCP tools

• Tracks retrieval telemetry and evaluation metrics

Architecture

User Query
    ->
Knowledge Fabric
    ->
Evidence Package
    ->
AI Assistant

Retrieval pipeline:

Sources
  ->
Ingestion
  ->
Chunking
  ->
Embeddings
  ->
Lexical + Vector Retrieval
  ->
Hybrid Fusion (RRF)
  ->
Evidence Package

Quick start

git clone <repository-url>
cd knowledge-fabric
python3 -m pip install -e ".[dev]"
docker compose up -d
python3 -m pytest

Run persistent ingestion:

knowledge-fabric-ingest --path sources --recursive --embed

Run MCP server:

knowledge-fabric-mcp

MCP tools

• retrieve_evidence

• get_document

• explain_retrieval

These tools return retrieval outputs and metadata, not final prose answers.

Documentation map

• docs/README.md

• docs/architecture.md

• docs/local-setup.md

• docs/evidence-package-contract.md

• docs/retrieval-evaluation.md

• docs/model-provider-strategy.md

• docs/security-and-data-boundaries.md

• docs/deployment-runbook.md

• docs/phase-1-knowledge-fabric/README.md

Current roadmap

1. Expand evaluation datasets and benchmarks.

2. Add optional reranking extension points.

3. Improve ingestion operational ergonomics for larger corpora.

4. Continue hardening local-to-production deployment guidance.

Contributing

Contributions are welcome. If you’re fixing a bug or adding retrieval capabilities, please include tests and docs updates in the same change whenever possible.

See CONTRIBUTING.md for full workflow details.

License

Apache License 2.0. See LICENSE.

Non-goals

• Workflow execution

• Approval orchestration

• Product-specific integrations

• Customer-specific deployment internals