This page’s configuration is based on FastGPT v4.17.0.
What this category of data looks like
Daily documents for this category include technical proposal books, customer requirement survey forms, quotation lists, and demo scripts. Data sources are customer interviews, bidding documents, vendor product materials, and historical project proposals. Update cadence is per order or weekly. Individual project documents are updated in real time per requirement changes, while bulk documents are synced weekly. Document structures include specialized technical terms, project boundary descriptions, quotation details, and demo workflows. Fields cover customer requirement points, technical routes, implementation scope, and cost allocation. There is no unified fixed format, and content must be customized based on current project needs.
What constraints do these characteristics impose on the knowledge base retrieval and recall link
Diverse document types with no unified format range from long-text technical proposal books to short-text quotation lists. Retrieval must adapt to content with varying lengths and structures. Each project binds data to exclusive data sources, so precise filtering of cross-project irrelevant content is required to avoid mixing unmatched proposal information during pre-sales Q&A. Flexible update frequency requires support for both incremental sync and full update modes to ensure recalled content aligns with current project status. A high proportion of specialized technical terms requires retrieval matching accuracy to meet business requirements, preventing errors in clarifying proposal boundaries due to low-relevance results.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Pre-sales proposals contain a large number of specialized technical terms. This threshold filters low-match irrelevant results while retaining content that meets required relevance standards |
limit | Top 8 entries | Pre-sales Q&A covers multiple scenarios such as proposal boundaries, quotations, and demos. Excessive results exceed context length limits. This quantity balances information completeness and retrieval efficiency |
searchMode | vector | Data primarily consists of specialized technical content. Vector retrieval is better suited for term matching than keyword matching, improving recall accuracy |
usingReRank | Enabled | Content in pre-sales proposals has strong relevance. Reranking optimizes the ranking position of high-relevance results, improving efficiency for users to obtain key information |
datasetSearchUsingExtensionQuery | Enabled | Customer questions may contain implicit technical requirements. Extended queries can match specialized content not directly referenced, improving the coverage of Q&A responses |
collectionIds | Bind corresponding vendor products and historical proposal collections per project | Pre-sales proposals are customized per project. Binding exclusive collections avoids retrieving cross-project irrelevant data, ensuring recalled content accurately matches the current project |
Three common misconfigurations
- Phenomenon: API calls to sync knowledge base collections prompt authentication failure. Cause: The dataset access key is not configured correctly, or the calling API account has not been granted access permissions for the corresponding knowledge base collection.
- Phenomenon: The number of returned retrieval results does not match the configured
limit. Cause:collectionIdsis not specified correctly, causing the retrieval scope to cover irrelevant knowledge base collections, and results are automatically filtered by the system. - Phenomenon: Knowledge base retrieval response takes too long. Cause: Both
usingReRankanddatasetSearchUsingExtensionQueryare enabled, andlimitis set too high, increasing computational load from retrieval, extended queries, and reranking.
How to confirm the configuration is correct
- Verify the
collectionIdsconfiguration, confirm that only exclusive data sources corresponding to the current pre-sales project are bound, such as vendor product materials and historical project proposals. - Initiate a retrieval test using simulated customer questions, check that the number of returned results matches the preset
limit. - Review retrieval logs to confirm that the trigger status of
usingReRankanddatasetSearchUsingExtensionQueryaligns with the configured requirements. - Test different types of question scenarios, such as technical route inquiries, quotation verification, demo script adjustments, confirm that the relevance of retrieval results meets business expectations.
The values provided are common starting points and should be measured 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.