peepshow/ sinks/ pgvector

Reel #17 Vector-in-SQL

peepshow sink / pgvector

pgvectorVector search on top of the postgres sink.

Keep runs + frames + embeddings in one Postgres. One database, full SQL over vectors.

drop · process · pgvector

What it does

Write both the structured archive (from the postgres sink) and frame embeddings into the same Postgres instance using the `pgvector` extension. One database for both metadata and RAG.

When to reach for it

  • Single source of truth: SQL + vector search in the same `SELECT`
  • Backup, replicate, and restore everything with one pg_dump
  • Avoid running three services (sql + object store + vector DB) — one Postgres covers it

Install

npm i -g peepshow
npm i -g pg

Use it

DATABASE_URL=postgres://... \
peepshow ./clip.mp4 --sink pgvector

Make it automatic

Register the sink once — every run fires it afterward. Scope by --when so it only runs for matching videos.

peepshow sinks add pgvector
peepshow sinks add pgvector --when extension=mp4,mov
peepshow sinks add pgvector --when path=/Volumes/Work/

Configuration

  • DATABASE_URL Postgres connection string. The `vector` extension must be installed. required
  • PEEPSHOW_PGVECTOR_DIM Embedding dimension. Default 1536 (OpenAI).

Use with an LLM agent

Every peepshow sink reads its config from env vars and receives a single JSON payload on stdin. An LLM agent (Claude Code, Cursor, Windsurf, Gemini, Codex) can drive the pgvector sink automatically when three things are true:

  • the env vars below are exported in the agent's shell (or a project .env it can load),
  • the peepshow CLI is on PATH — install with npm i -g peepshow,
  • a peepshow auto-sink is registered for the run (optional but recommended — makes invocation zero-argument).

1. Set the environment

# Add to ~/.zshrc, ~/.bashrc, or a project .env the agent can load
export DATABASE_URL="..."

2. Register as an auto-sink

peepshow sinks add pgvector
peepshow sinks add pgvector --when extension=mp4,mov

3. Example LLM session

You → drop a .mov into Claude Code.

Claude → auto-invokes /peepshow:slides ./clip.mov. peepshow extracts frames + audio, the pgvector sink forwards the run to the configured collection. Claude replies with a summary and a link to the created record.

The transcript text is a natural embedding target — index it next to the frames.

Write your own

A sink is any executable that reads the --emit json payload on stdin. Shell, Node, Python, Go — the spec's in docs/PLUGINS.md. Register persistent ones with peepshow sinks add-cmd 'your-command'.