Search every word you've ever heard
PodPast.ai indexes every podcast episode and YouTube video you add as a semantic vector, not a keyword index. Ask a question in plain English and the system retrieves the most relevant transcript passages from across your entire library — regardless of whether your exact words appear in the transcript.
How semantic search works
Add any podcast RSS feed or YouTube URL to your Pod
PodPast.ai fetches the complete episode catalogue. There is no source limit — add as many feeds as you consume.
PodPast.ai transcribes and embeds every episode into a 256-dimensional vector index
Each transcript chunk — roughly 150–200 words — is embedded as a vector. Chunks retain timestamp, episode, feed, and speaker metadata. Free transcript sources (YouTube captions, RSS tags) are used first; Deepgram nova-2 handles audio-only episodes.
Ask any question in Claude or via the search API — results rank by semantic similarity
The query is embedded with the same model, and pgvector finds the nearest neighbours. A re-ranking pass weights by keyword overlap and recency. Every result includes a timestamp link to the source.
Semantic vs. keyword vs. per-episode search
| Capability | PodPast.ai | Keyword search | Snipd (per-episode) |
|---|---|---|---|
| Finds conceptually related passages | ✓ | ✗ | ✗ |
| Cross-corpus (all feeds) | ✓ | Depends | ✗ Per-episode only |
| Handles synonym variation | ✓ | ✗ | ✗ |
| Timestamp citations | ✓ | Varies | ✓ Per highlight |
| No source limit | ✓ | Varies | ✓ |
| Works inside Claude | ✓ via MCP | ✗ | ✗ |
Semantic search questions
- What does 'semantic search' mean for podcasts?
- Semantic search retrieves results based on meaning, not exact word matches. If you ask 'what did they say about interest rates affecting startup valuations', PodPast.ai finds relevant transcript passages even if none of them use those exact words — it matches concepts like 'monetary policy', 'cost of capital', and 'down rounds' because they are semantically related. This is implemented via a vector embedding model that maps text to a 256-dimensional space, where proximity equals conceptual similarity.
- How accurate is the transcript search?
- Accuracy depends on transcript source. YouTube captions are timestamp-accurate to within 1–2 seconds and have error rates below 5% for standard English speech. Deepgram nova-2, used for audio-only episodes, achieves 85–92% word accuracy on conversational speech. RSS transcripts vary by podcast. PodPast.ai displays the transcript source on each search result so you know which accuracy tier you're working with.
- Can I search across multiple podcasts at once?
- Yes — cross-corpus search is the primary function. A single query searches your entire vault by default. You can optionally scope a search to a specific feed or date range, but the default is to search everything. This is what distinguishes PodPast.ai from per-episode tools like Snipd: you get a unified index across every feed you've added.
- What languages does PodPast.ai support?
- Semantic search works in any language that Deepgram nova-2 supports, which includes English, Spanish, French, German, Portuguese, Italian, Dutch, Hindi, Japanese, and Korean. YouTube captions work in any language YouTube provides. Cross-language semantic search (querying in English, retrieving results in Spanish) is on the roadmap but not yet available.