Semantic Search¶
Semantic Search builds a reusable local semantic index from uploaded documents and retrieves the most similar chunks for a query.
Index Builder¶
The index builder supports:
- PDF/TXT/DOCX upload;
- Chunk Size;
- Chunk Overlap;
- Clear Existing Index;
- Append to Existing Index;
- embedding diagnostics.
Embeddings are generated locally through sentence-transformers.
Document identity¶
The embeddings table retains source-document identity in addition to chunk text and vector data.
This enables document-aware query behavior.
Semantic Query¶
Controls include:
- Top K;
- Minimum Similarity;
- Group by Document;
- query text.
Group by Document¶
Group by Document is functional rather than decorative.
When enabled:
- chunks are scored by cosine similarity;
- results are ordered by similarity;
- only the strongest result for each source document is retained;
- Top K is applied to the document-grouped result set.
Result routing¶
Selected semantic chunks can be:
- sent to Text Generation;
- sent to Document Q&A;
- saved into shared prompt/document context.
Failure behavior¶
Missing embedding support or malformed/stale vector dimensions degrade through guarded paths rather than producing an uncaught UI-to-model exception.