Skip to main content

ADR 006: Document Lineage via Qdrant Payload Aggregation

Context​

Users need to browse ingested content at the document level (source → chunks). Adding a dedicated SQL store for document metadata was considered but adds operational complexity.

Decision​

Store all document metadata in the existing Qdrant doc.* payload fields (source_id, source_type, filename, mime_type, ingest_date). Aggregate document summaries at API time by scrolling and grouping chunks by doc.source_id.

Consequences​

  • No additional database dependency.
  • Aggregation happens at query time — acceptable for collections with fewer than 10K chunks.
  • Document-level operations (delete all chunks for a document) are straightforward with Qdrant filters.

Alternatives considered​

  • Separate PostgreSQL table for documents — adds operational burden.
  • Redis cache of document summaries — stale data risk without invalidation logic.