What this category of data looks like
Typical data for this category comes from financial systems, project management systems, contract ledgers, and business department requests. Daily documents include annual budget sheets, project proposal documents, project portfolio dashboards, and investment review reports. Fields cover budget amounts, project cycles, priority levels, investment returns, and more. Units include yuan, ten thousand yuan, days, and months. Data update frequency is mostly monthly, quarterly, or annually. Single document volume grows with the number of projects. Some dashboard documents contain multi-dimensional tables and value comparison content.
What constraints do these characteristics impose on the knowledge base retrieval and recall link
Differences in data formats from multiple sources require retrieval to support both structured tables and unstructured text, to avoid missing critical numeric information. Low-frequency updated but high-value documents need a stable recall mechanism to ensure consistent results. Fields such as budget amounts and project priorities in documents require retrieval results to be associated with corresponding metadata, to ensure answers can trace back to specific sources. At the same time, the number of retrieval results for this category must be controlled, to avoid too many low-relevance content interfering with decision-making references for budget priority sorting and project review.
How to set the configuration
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the mixed numeric and text features of budget-related documents, filters out low-relevance retrieval results |
limit | Top 8–12 results | Adapts to the multi-document association needs of project portfolio management, avoids excessive single-round recall causing context overflow |
searchMode | hybrid | Combines keyword and semantic retrieval, supports exact numbers in budget sheets and text descriptions in project proposals |
usingReRank | Enabled, take top 3–5 results | Prioritizes results strongly related to budget priorities and project reviews, improves answer accuracy |
datasetSearchUsingExtensionQuery | Enabled | Automatically supplement business-related extended query terms such as budgets and project cycles, improving recall accuracy |
collectionIds | Bound by business module | Only retrieve dedicated datasets for budget management and project management, exclude irrelevant business documents |
Three common mistakes
- Phenomenon: Chinese file names uploaded via API display garbled characters, or the file name field returned by the interface is empty. Cause: The correct UTF-8 encoding header was not configured in the API request, resulting in failed file name parsing.
- Phenomenon: Knowledge base retrieval and AI chat take too long, and page loading is stuck. Cause: The number of recalled entries exceeds a reasonable range, and the
usingReRankreranking function is not enabled, resulting in a large number of low-relevance documents being loaded. - Phenomenon: AI stops mid-way when answering based on the knowledge base, without completing the full response. Cause: The total token count of single-round recalled documents exceeds the model's context limit, and the number of recalled results is not controlled via the
limitparameter.
How to confirm the configuration is correct
- A budget sheet file with a Chinese file name is uploaded, and the file name in the knowledge base is verified to display normally without garbled characters.
- A test query for budget priority is initiated, and the number of returned retrieval results is checked to match the
limitvalue range set in the configuration. - Retrieval results are checked for metadata fields such as project numbers and budget amounts, confirming that the metadata association function operates correctly.
- More than 10 consecutive retrieval requests are sent, and response latency is checked to stay within a reasonable range.
The values provided here 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.