# Data Pipeline

Source data and documentation for the vector index.

## Files

| File | Description |
|------|-------------|
| `episodes_website.json` | Episode metadata (108 episodes): titles, guests, YouTube URLs |
| `episodes_embedding_chunks.csv` | 14,865 transcript chunks with timestamps |

## What Gets Embedded

Each chunk is a structured text block that includes episode context and transcript content:

```
Episode: 455
Chapter: Introduction

BEFORE (context):
[~50 words from previous segment]

MAIN (target):
Adam Frank: If we don't ask how long they last, but instead ask 
what's the probability that there have been any civilizations at all...
[~300 words - the core content we're searching]

AFTER (context):
[~50 words from next segment]
```

**Key design choices:**
- Episode number and chapter title are embedded with the text — this helps retrieval when users ask about specific episodes or topics
- The MAIN section contains the actual transcript with speaker names inline
- BEFORE/AFTER provide context so the LLM understands the quote in its original setting
- Overlapping windows ensure content isn't awkwardly split at chunk boundaries

## Vector Store

**Embeddings:** OpenAI `text-embedding-3-small` (1536 dimensions)

**Index:** FAISS flat index (`IndexFlatL2`)
- Simple brute-force similarity search
- Works well for our scale (~15K vectors)
- No approximation — returns exact nearest neighbors
- Trade-off: Larger indices (100K+) would benefit from IVF or HNSW for speed

**Storage:** `../faiss_index.db/`
- `index.faiss` — vector data
- `index.pkl` — metadata (episode, timestamp, URL, chapter)

## Adding New Episodes

To add new episodes or rebuild the index:

1. Scrape new episode metadata and transcripts
2. Process transcripts into chunks using the same BEFORE/MAIN/AFTER format
3. Append to `episodes_embedding_chunks.csv`
4. Rebuild the FAISS index:
   ```bash
   uv run python build_index.py
   ```
5. Check index status:
   ```bash
   uv run python build_index.py --check
   ```

The flat index must be fully rebuilt — it doesn't support incremental additions. For a production system with frequent updates, consider using a vector database like Pinecone or Weaviate that supports upserts.

**Current data (as of Feb 2026):** 108 episodes (#276–#491), newest: Peter Steinberger (OpenClaw)

## Privacy

No user data is stored here. All content is from publicly available podcast episodes.