Feature Idea: Metadata filtering for retrieval (pre-filter chunks/entities by tags, file path, etc.)

[hata33]

Problem

Currently LightRAG has no metadata filtering in the retrieval path. QueryParam has no filter field, and all vdb.query() calls pass only query + top_k + query_embedding. Users cannot restrict retrieval to specific documents, file types, date ranges, or custom tags.

Additionally, chunks carry very sparse metadata (TextChunkSchema only has tokens, content, full_doc_id, chunk_order_index) — no file_path, source, or arbitrary user metadata is stored as a queryable field on vector records.

Proposed Feature

1. Persist metadata on chunks/entities/relations — extend the ingestion pipeline to store user-provided metadata (e.g. file_path, doc_type, tags, date) as payload fields in the vector DB.

2. Add a metadata_filter parameter to QueryParam — allow users to pass filter conditions (e.g. {"file_path": {"$contains": "2024"}} or {"tags": {"$in": ["finance"]}}).

3. Thread the filter through vdb.query() — native backends (Qdrant Filter, Milvus expr, PostgreSQL WHERE) already support filtering; in-memory/FAISS backends would use post-filtering.

Use Cases

• Multi-tenant: restrict retrieval to a specific user or workspace subset
• Time-scoped queries: only search documents from a specific date range
• Document-type filtering: search only PDFs, only meeting notes, etc.
• Custom tags: user-defined labels on documents at ingestion time

Current Workaround

None — users must maintain separate LightRAG instances per filter dimension, which is wasteful.

Question

Is this on the roadmap? Would a PR adding metadata filtering (starting with Qdrant + PostgreSQL backends) be welcome?