Vector Search Options

Select a model compatible with your LLM provider (such as OpenAI or Azure OpenAI). The choice of model affects the quality and language support of semantic search.

• Use a model optimized for embeddings (for example, text-embedding-ada-002 for OpenAI).
• Ensure the model supports the languages present in your documents.
• Changing the model may require reindexing to update stored vectors.

The selected model determines the dimensionality and semantic accuracy of the generated vectors, directly affecting retrieval quality.

When enabled, Grooper segments document content into overlapping 8K-token chunks and stores them in a dedicated chunked vector index. This allows for passage-level retrieval and overcomes the context window limitations of most embedding models.

• Each chunk is embedded separately, enabling retrieval of the most relevant passage within a long document.
• Both a main (full-document) and a chunked index are maintained in the search service.
• When disabled, only the first N tokens (for example, 8,192) are embedded per document, and long content is truncated.

Recommended for documents that frequently exceed the model's context window (such as long contracts or reports). Improves recall and accuracy for RAG and semantic search scenarios.

Note: Enabling this option increases storage and indexing time, as two indexes are maintained.

When enabled, metadata fields (such as document type, date, or custom fields) are stored alongside content vectors in the index. This allows for more granular filtering and faceted search in semantic queries.

• Filter semantic search results by document type, date range, or other metadata.
• Combine keyword and semantic filters for precise retrieval.

Best practices:

• Include only metadata fields that are relevant for filtering or grouping search results.
• Review privacy and compliance requirements before indexing sensitive metadata.