matrx-rag 0.1.73

Multi-tenant RAG: hybrid retrieval, chunking, embeddings, ingestion, PDF + image + repo pipelines, agent-extraction indexing, priority-aware ranking.

matrx-rag

Multi-tenant RAG: hybrid retrieval, chunking, embeddings, ingestion, PDF + image + repo pipelines, agent-extraction indexing, NER/KG writes, priority-aware ranking — extracted from aidream/services/rag/ so any host (aidream's cloud server, matrx-local) can consume the same stack.

Documentation hub: docs/rag_and_ner/README.md
(Vision in docs/knowledge/ · code truth in docs/rag_and_ner/reality/ · backlog in 00_CLEANUP.md.)

What it does

• Ingestion: ingest_source(source_kind, source_id, ...) — chunk → embed → upsert; NER inline for non-code sources.
• PDF / image / code repo pipelines with provenance.
• Hybrid search: pgvector HNSW + lexical FTS → RRF → optional Cohere → MMR.
• Priority-aware ranking: agent-extracted chunks with non-zero priority.
• Agent-extraction → RAG: extraction_indexing materializes page-extraction payloads as chunks.
• Data stores: curated buckets with scoped search.
• Eval harness: retrieval + answer quality assessment.

Host integration

• aidream wires the package in aidream/package_integration.py::_configure_matrx_rag (NER extractor, embedding cache, ORM models, search callable).
• aidream/services/rag/ is a re-export shim; HTTP routes live in aidream/api/routers/rag.py.
• Auto-ingest (gates, budget, scope suggestions) lives in aidream/services/auto_ingest/ — not in this package.

Schema

Migrations: matrx_rag/migrations/ (mirrored in db/migrations/ until package owns the runner).
ORM / pgvector notes: docs/rag_and_ner/reality/04_ORM_AND_SCHEMA.md.

Package development

See CLAUDE.md for injection seams, test baselines, and the no-aidream-imports rule.