Step 3. Retrieval Test and Recall Strategies
Configure Retrieval and Recall Strategies
Recall Settings
Once the knowledge base is linked to the prompt, character, or workflow, you can configure retrieval and recall settings:
Mixed recall: Comprehensive results returned using vector and full-text recall method.
Vector recall: Using vectors for text correlation queries, suitable for cross-language and meaning-based queries.
Full-Text recall: Use traditional full-text search, suitable for finding specific keywords, used for terms, IDs, or abbreviations.
GraphRAG only:
Local recall: Return search results using related nodes and nearby neighbors
Global recall: Use all nodes, filter and summarize the search results and return them (higher cost)
Recall parameters:
Parameters
Description
Recall methods
Token
Select how many tokens to return from the retrieval results. The larger the number, the longer results are returned.
All
Correlation
The threshold of the results to be returned based on the correlation in semantics.This configuration can filter out some low-relevance search results.
Mixed and Vector recall only
Re-rank Model (not supported in GraphRAG)
The re-rank model assesses the semantic alignment between the user's query and a set of potential results, repositioning them according to their semantic compatibility to enhance the outcome of semantic ordering. It operates by calculating a relevance score for each document in relation to the user query and produces a list of documents organized by decreasing relevance. iSiri currently supports the following re-rank models: Cohere rerank multilingual v2.0, bge-reranker-large, bge-reranker-small.
Last updated