Features

Vector Search (Semantic Retrieval)

File References

• supabase/functions/search_chunks/index.ts
• src/utils/transcription/TranscriptionService.ts (for calling the search function)

What It Does

The vector search system:

1. Takes user input
2. Uses GPT to parse into keywords & topics
3. Retrieves relevant chunk IDs by comparing embeddings
4. Sorts chunk results by similarity
5. Optionally merges results with GPT for final summary

    
flowchart TD
    A[User Input] --> B[GPT Processing]
    B --> C[Keywords]
    B --> D[Topics]
    C --> E[Vector Embeddings]
    D --> E
    E --> F[Compare with Chunk Embeddings]
    F --> G[Sort by Similarity]
    G --> H[Relevant Chunks]
    H --> I{Merge with GPT?}
    I -->|Yes| J[Final Summary]
    I -->|No| K[Raw Results]

  

Key Points

• Query is broken down into "top_3_keywords" and "top_3_topics"
• Chunk embeddings stored in pgvector columns for fast KNN lookups
• Thresholds/tuning (similarity) can be adjusted in the Edge Function

Document Processing / Chunking

File References

• supabase/functions/insert_transcript/index.ts
• supabase/functions/process_chunk/index.ts
• supabase/functions/_shared/chunks.ts
• src/utils/audio/ (for capturing audio, if using live mode)

Process Flow

1. Transcript Ingestion: User either pastes a transcript or records one live
2. Clean & Split: Removes filler words, splits into sentences
3. DetectTopicChunks: Uses sliding window + cosine similarity threshold for chunk boundaries
4. Process & Store:
   • Each chunk sent to process_chunk Edge Function
   • GPT-based metadata extraction (keywords, topics, sentiment, summary)
   • Storage in PostgreSQL with embeddings

    
flowchart TD
    A[Transcript Input] -->|Paste or Record| B[Transcript Ingestion]
    B --> C[Clean Text]
    C --> D[Split into Sentences]
    D --> E[Detect Topic Chunks]
    E -->|Sliding Window| F[Chunk Boundaries]
    F --> G[Process Each Chunk]
    G --> H[GPT Metadata Extraction]
    G --> I[Generate Embeddings]
    H --> J[Store in PostgreSQL]
    I --> J
    
    subgraph Metadata
        H --> K[Keywords]
        H --> L[Topics]
        H --> M[Sentiment]
        H --> N[Summary]
    end

  

Query Understanding

File References

• src/components/Search/SearchHeader.tsx
• src/pages/SearchResults.tsx
• supabase/functions/search_chunks/index.ts

Result Ranking & Summaries

File References

• src/pages/SearchResults.tsx
• supabase/functions/search_chunks/index.ts

Implementation

• Chunks sorted by vector similarity
• GPT used for final "ragPrompt" to produce user-friendly summary
• Expandable chunks with highlighted relevant lines

Note: Threshold tuning is crucial - too high might miss context, too low may include noise.