Embedding

Encodes chunks into vectors for semantic search & hybrid ranking.

Embedder

• Default: lightweight sentence transformer (config: EMBED_MODEL).
• Batch size & concurrency tuned to avoid memory spikes.

Failure & Retries

• If the model fails, the chunk is marked pending; the scheduler retries later.
• We never block ingestion of the catalog on embedding failures.

Stored Fields

• entity_uid
• chunk_id
• vector (list of floats / DB-specific binary)
• dim, model_id, created_at

Replacement Policy

• Re-embedding occurs only when:
• Chunk text or weight changed, or
• Embedder model_id changed, or
• Admin forced re-embed.