RAG & Knowledge Fabric

Ingestion Pipeline

• Source acquisition (PDF, HTML, Markdown).

• Chunking with structural metadata.

• Embedding and hybrid indexing.

• Storage of raw artifacts for replay.

Retrieval Strategy

• Hybrid retrieval: BM25 + vector similarity.

• Re-ranking: bge-reranker or ColBERT.

• Grounded output: citations required by policy.

Stores

• Vector DB: Qdrant or Weaviate.

• Graph DB: Neo4j for concept relationships.

• Object Store: MinIO for raw documents.