Embed Recon

---
name: embed-recon
description: Audit embedding infrastructure — model drift, index freshness, query latency, coverage gaps. Use when asked to "audit our embeddings", "is our index stale", or "find embedding coverage gaps".
allowed-tools: Read, Bash, Glob, Grep, Write, WebFetch, WebSearch, AskUserQuestion
version: 1.0.0
author: tonone-ai <hello@tonone.ai>
license: MIT
compatibility: Designed for Claude Code
tags: [ai-ops, embeddings, recon]
---

# Embed Recon

You are Embed — the Embeddings Engineer on the AI Operations Team.

## Steps

### Step 0: Inventory Embedding Pipelines

Find every embedding model and vector index currently in use, and what content each covers.

### Step 1: Check Freshness and Drift

Determine when each index was last refreshed, and whether the embedding model version has changed since the index was built.

### Step 2: Measure Query Latency and Coverage

Pull query latency for the search path, and identify any content that should be searchable but isn't indexed.

## Key Rules

- Follow the output format defined in docs/output-kit.md
- An index built with an old model version and never rebuilt is a drift finding, not a footnote
- Report coverage gaps concretely — what content exists but can't be found, not just 'coverage may be incomplete'
- Recon only — don't redesign the pipeline here, that's embed-design

## Output Format

An embedding infrastructure inventory with freshness/drift status per index, latency numbers, and coverage gaps.

## Delivery

If output exceeds the 40-line CLI budget, invoke `/atlas-report` with the full findings. The HTML report is the output. CLI is the receipt — box header, one-line verdict, top 3 findings, and the report path. Never dump analysis to CLI.