sai-podcast-rag-index

name

SAI-podcast-rag-index

description

Atomic SAI skill. Index a directory of transcript .md files (YAML frontmatter, one per episode) into an existing ChromaDB RAG store by delegating to a rag-index-updater-style index.py. Pure mechanism — no operator paths or collection baked in; the caller supplies indexer_scripts / manifest_path directly or via a private config_path. Hash-skips unchanged files. Use whenever you need to index transcripts, add transcripts to a RAG, or embed episodes for semantic lookup.

Trigger manifest — podcast-rag-index (atomic, autonomous)

Code: skills/podcast-rag-index/ Manifest: skill.yaml (3-tier cascade) Entry point: runner.py → run(inputs)

Cascade plan

1. locate_indexer   (rules) → resolve indexer_scripts (input or config_path); fail-closed if unset
2. scan_inputs      (rules) → diff input_dir vs manifest (new+changed only)
3. run_indexer      (rules) → shell to index.py → ChromaDB upsert

Invocation

.venv/bin/python3.12 \
  skills/podcast-rag-index/runner.py run \
  --inputs-json '{
    "input_dir":   "~/Media/transcripts/<show-slug>",
    "config_path": "<your-private-overlay>/config/podcast_ingest.yaml"
  }'

config_path supplies operator values (indexer_scripts, manifest_path) from the private side. Or pass indexer_scripts directly. "force": true re-indexes even if hash unchanged. Invoke by absolute path (hyphenated folder name).

Why a thin wrapper

The operator's index.py owns embed + chunk + upsert + the ChromaDB collection. This skill only scopes the scan to one directory and reuses that indexer — one source of truth, no second codepath to drift.

Composes with

• Upstream: SAI-local-transcribe
• Companion: a general-purpose rag-index-updater (full-tree variant)
• Downstream consumer: a rag-content-lookup (semantic search over results)