// septim forge tools · db / sql

pgvector Query Builder

client-side · zero network calls · no telemetry · no login

// build query

// presets

Table name

Vector column

Dimension

Distance metric

Result limit (k)

Query vector source

WHERE filter (optional)

SELECT columns (comma-separated, or *)

Also generate CREATE INDEX DDL

// similarity search query

// index DDL

Fill in the fields above and click "Build query" to generate SQL.

// which distance metric to pick

Three operators. Three use cases.

Every pgvector distance metric maps to a different mathematical notion of similarity. Picking the wrong one silently returns valid-looking but semantically wrong results.

Metric	Operator	Best for	Avoid when
Cosine distance	<=>	Normalized embeddings (OpenAI, Cohere, most sentence-transformers). The default for RAG. Measures angle, ignores magnitude.	Embeddings are not normalized to unit length.
L2 distance	<->	Raw un-normalized vectors where magnitude carries meaning. Some image embeddings trained on L2.	Typical text embeddings — cosine is almost always better.
Inner product	<#>	Models explicitly trained with dot-product similarity. Note: pgvector sorts ASC — invert sign for max similarity.	Generic embedding models without a known training objective.

Quick rule: OpenAI text-embedding-3-*, Cohere Embed, or any standard sentence-transformers model — use cosine (<=>).

// index selection

HNSW vs IVFFlat

pgvector ships two approximate nearest-neighbor index types with different operational tradeoffs.

Factor	HNSW	IVFFlat
Build time	Slower (builds the graph)	Faster (builds the inverted file)
Query speed	Faster at query time	Slower, especially at high recall
Memory	Higher — stores graph edges	Lower baseline
Recall	Higher out of the box	Requires tuning `lists` and `probes`
Incremental inserts	Handles well	Degrades; rebuild periodically
Recommended for	Most production RAG applications	Constrained memory, large static datasets

Start with HNSW, m=16, ef_construction=64. Tune ef_search at query time for recall. Set SET hnsw.ef_search = 100; in the session before querying.

// verify zero network calls

How to confirm this tool never phones home

// devtools network tab

Open DevTools (F12 or Cmd+Option+I), click the Network tab, then use the query builder above. After the page finishes loading its fonts and CSS, build a query and watch the Network tab. Zero new requests fire. All computation runs in the browser JavaScript engine — no data leaves your machine.

// building a production AI application on postgres?

Septim Drills includes 25 Claude Code production skills covering SQL migrations, query optimization, and pgvector integration patterns. Drop into ~/.claude/skills/, reference by name, ship faster.

Septim Drills — $29 once →