AI & Vector Search

This page outlines an embedding-based search pipeline with pgvector.

Schema

CREATE TABLE IF NOT EXISTS post_embeddings (
  post_id UUID PRIMARY KEY,
  embedding vector(1536),
  meta jsonb
);

Pipeline

• Chunk & normalize content (strip HTML, split by tokens, limit length).
• Generate embeddings (model choice, batch sizing, retry/backoff).
• Store: upsert into post_embeddings with metadata (lang, tags, ts).
• Query: accept a query string → embed → cosine distance search.

Examples

-- Top 10 most similar posts
SELECT post_id
FROM post_embeddings
ORDER BY embedding <-> $1
LIMIT 10;

GraphQL integration

• Expose similarTo(text: String!, limit: Int) resolving to the SQL above.
• Combine filters: intersect vector result with domain filters (tenant, tags).

Jobs & backfills

• Use a projector or background job to (re)embed changed entities.
• Track last embedded version to avoid redundant work.

Privacy & safety

• Redact PII before embedding; avoid leaking secrets in prompts.
• Store minimal metadata; encrypt at rest if required by policy.

Cost control

• Batch embeddings; cache by content hash; schedule low-priority backfills.

See also

• Vision → Edge & AI (direction): /vision