Overview of FastGPT Dataset Search
FastGPT dataset search does not operate as a simple direct user question-to-vector database lookup workflow. Instead, it integrates multiple complementary retrieval and refinement components to generate high-quality cited content for large language model processing. The pipeline combines query rewriting, multi-channel candidate recall, result merging, secondary sorting, and final filtering to adapt to diverse input types and search requirements.
Step-by-Step Search Pipeline
The full search workflow follows six standardized stages as defined by the FastGPT dataset engine:
- Query Rewriting: Resolves coreferential references in multi-turn conversations and expands the original query to improve semantic richness and search relevance for complex user requests.
- Candidate Recall: Selects one or more recall channels based on configured parameters, including semantic search, full-text search, or hybrid search, to retrieve a preliminary set of candidate content.
- Image-Aware Search: If the input payload includes images, the pipeline uses either image description search or image vector search, dependent on the capabilities of the deployed computer vision model.
- Result Merging: Applies Reciprocal Rank Fusion (RRF) to combine results from all active search channels into a single unified candidate set, eliminating duplicate or low-priority entries early in the workflow.
- Secondary Sorting: Runs a reranking step to reorder the merged candidate list, refining the relevance of the top candidates before final filtering.
- Final Filtering: Applies similarity filtering and citation limits to trim the candidate set, producing the final curated cited content that is passed to the connected large language model.
Configurable Search Parameters
Users can adjust the following parameters to tailor the search pipeline to their use case, all defined within the FastGPT dataset engine configuration:
| Parameter | Valid Options | Description |
|---|---|---|
| Search Mode | Semantic Search, Full-Text Search, Hybrid Search | Specifies the primary candidate recall method(s) |
| Enable Query Rewriting | true, false | Toggles the coreference resolution and query expansion stage |
| Enable Image Search | true, false | Activates image-based search processing for image inputs |
| Enable Reranking | true, false | Toggles the secondary relevance sorting step |
| Similarity Threshold | Float between 0 and 1 | Filters out candidates with a similarity score below the specified threshold |
| Maximum Citation Count | Positive integer | Limits the total number of cited content items included in the final output |
Source: FastGPT official source
Applicability and version scope
Use this page for the documented Tutorials scenario. Confirm the FastGPT, dependency, API, and deployment versions in the official source before applying a change.
Safety guardrails
Use [REDACTED_CREDENTIAL] for credentials and private data. Confirm the documented environment and version before review.
Rollback guidance
Restore the prior technical-content authority snapshot. Restore saved configuration and data snapshots, then repeat the smallest verification scenario.