This page is compatible with FastGPT v4.17.0.
What this category of data looks like
This type of data mainly comes from customer interviews, bidding documents, vendor product materials, and historical project plans, with an update cycle of per order or weekly. Daily produced documents include technical proposal books, customer requirement survey forms, quotation lists, and demo scripts. Technical proposal books are structured into modules such as technical roadmap, implementation boundaries, and quotation configuration. Quotation lists include individual service amount and total quotation fields. Requirement survey forms record customer functional and budget requirements.
What constraints these characteristics impose on the "re-ranking and hybrid retrieval" link
Since data sources include bidding-related documents, content must strictly match customer requirement boundaries. During hybrid retrieval, historical plans and vendor materials matching current pre-sales requirements must be recalled first. The update frequency is high-frequency adjustments within a single project cycle. Retrieval results must cover the latest vendor product parameters and the customer's latest requirements. Daily documents have large structural differences: long texts in technical proposal books and structured data in quotation lists need to be adapted to retrieval logic separately, to avoid irrelevant content interfering with core matching results, while ensuring retrieval results can directly support proposal boundary design and quotation verification.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
usingReRank | Enabled | Pre-sales proposals require strict matching of requirement boundaries, and re-ranking can filter low-relevance historical plans and vendor materials |
similarity | 0.72–0.85 | Meets the strict matching requirements of bidding documents, and avoids recalling irrelevant non-pre-sales content |
limit | Top 10 entries | Balances retrieval coverage and result processing efficiency, and adapts to the document volume of a single pre-sales project |
searchMode | hybrid | Combines keyword retrieval and semantic retrieval, and adapts to mixed pre-sales document types with both structured and unstructured content |
| Re-ranking model call timeout threshold | 15 seconds | Prevents delays in model response from affecting rapid generation and verification of pre-sales proposals |
Three common mistakes
- Re-ranking model tests return 401 Unauthorized error. This occurs when the API key for the re-ranking model is not configured correctly, or the key permissions do not cover the re-ranking model call scope.
- When the re-ranking function is not enabled, the relevance metrics of retrieval results have no clear reference value. Basic semantic retrieval only relies on text semantic matching, and does not perform targeted sorting based on query intent, and cannot filter redundant content irrelevant to current pre-sales requirements.
- The number of results returned by hybrid retrieval exceeds the expected range. This happens when the value of the
limitparameter is not set correctly, resulting in too many low-relevance documents being recalled, which increases verification costs for pre-sales proposal boundary design.
How to confirm the configuration is complete
- Enter the FastGPT retrieval configuration page, and confirm that the
usingReRankswitch is in the enabled state. - Submit a test query related to pre-sales requirements, and check whether the returned results are sorted to prioritize matching the core requirements of the current pre-sales proposal.
- Check the re-ranking model call logs, and confirm that there are no 401 Unauthorized errors or timeout errors.
- Adjust the value of the
similarityparameter, and verify that the recall range of retrieval results adjusts reasonably with changes in the parameter.
The values provided here are common starting points, and should be evaluated against the reader's own samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-10-19.