What this category of data looks like
The data used for survey and design declaration and file review comes primarily from project survey reports, construction drawing design documents, approval submission files, and archived past qualified files. Data updates trigger alongside project submission progress. Single-project document updates have no fixed cycle. Batch archived data synchronizes according to completed project batches. Documents include both structured fields and unstructured content: Structured fields include project number, survey point coordinates, design load unit (kN/㎡), survey depth unit (meter), and similar items. Unstructured content includes PDF-format drawing descriptions and Word-format submission explanation documents.
What constraints these characteristics impose on knowledge base retrieval and recall
The mixed structured and unstructured nature of survey and design data means retrieval must cover both professional term matching and semantic association. This avoids missing key content with a single retrieval mode. The fact that data updates along with project progress requires knowledge base incremental synchronization to trigger per project batch. This prevents retrieval delays caused by full refreshes. Fixed units for professional fields require retrieval to retain unit-linked matching. This stops recall results with incorrect units from interfering with review judgments. The fixed archive format of past qualified files requires retrieval to match structured fields such as file number and approval submission batch. This improves the targeting of recall results.
How to set configurations
These configurations are formulated for the knowledge base retrieval feature of FastGPT v4.17.0:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.80 | Survey and design documents contain a large number of professional terms. This threshold range can filter irrelevant general specification documents while retaining approximate qualified file content |
limit | Top 8 entries | Declaration material review requires comparing multiple past qualified files. 8 results cover common submission scenarios and do not exceed context window limits |
searchMode | semantic+keyword | Survey and design documents have both structured fields and unstructured descriptions. The hybrid retrieval mode can match both keywords and semantically associated content |
usingReRank | Enabled | There are many professional terms in survey and design documents. Reranking can filter low-quality results that are semantically similar but do not match terms, improving the professionalism of recall results |
datasetSearchUsingExtensionQuery | Enabled | The wording of declaration materials may differ from the terms stored in the knowledge base. Extended queries can supplement synonym matching, for example, "load" matching "load" |
collectionIds | Enter the collection ID of the survey and design-specific knowledge base | Avoid retrieving documents from other categories, ensuring recall results only cover content related to survey and design |
Three Common Mistakes
- Phenomenon: Semantic retrieval fails to recall knowledge blocks with specific professional keywords. Cause:
datasetSearchUsingExtensionQueryis not enabled, and synonyms of professional terms are not covered, leading to an overly narrow matching scope. - Phenomenon: When using a strict question-and-answer template, retrieving image addresses prompts that no answer was found. Other templates respond normally. Cause: The strict question-and-answer template only matches structured text fragments, and does not adapt to image-based knowledge base content. General templates automatically recognize unstructured content and generate corresponding responses.
- Phenomenon: An API call returns an incorrect knowledge base ID error. Cause: The
collectionIdsparameter is not configured correctly, or the passed collection ID does not belong to the survey and design-specific knowledge base, leading to a mismatch in the retrieval scope.
How to Verify Correct Configuration
- Submit a simulated declaration material retrieval request. Confirm the returned results only include documents from the survey and design category, with no irrelevant content from other categories.
- Enter a query containing professional terms. Confirm the similarity scores of the recall results fall within the 0.72–0.80 range, and reranked results follow professional logic in their ranking.
- Call the API with the specified
collectionIdsparameter. Confirm the number of returned results matches the value set by thelimitparameter. - Test knowledge base content containing image addresses. Confirm retrieval behavior meets expectations across all template types.
The values given 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 2026-09-28.