What this category of data looks like
The application materials and case review data for municipal water utilities primarily originate from full-process documents of water projects. These include water intake permit applications, drainage permit submission materials, completion acceptance cases for pipeline network renovation projects, weekly water quality test reports, and more. Data sources cover scanned offline paper documents, files exported from online approval systems, and structured records entered into operation and maintenance systems. Update frequency follows project progress: individual project materials are updated multiple times between submission and approval completion, while daily operation and maintenance cases are updated in batches each month. Document structures include structured fields such as project number, submitting unit, and approval status, alongside unstructured body text. Field units use professional water industry units including cubic meters, megapascals, milligrams per liter, and others.
What constraints these characteristics impose on knowledge base retrieval and recall
There are many structured fields and professional terms in water utility application materials and review cases. Precise matching of key information such as project numbers and approval status is required to avoid retrieving irrelevant content. The length of individual case documents varies widely, ranging from a few hundred characters to tens of thousands of characters. When chunking documents, both completeness and relevance must be balanced. Materials from different projects must be isolated by project folders to prevent interference from cross-project retrieval. Additionally, the review scenario requires retrieval results to strictly match official document content in the knowledge base, with no extended generated non-original content allowed. These characteristics mean retrieval configurations must support precise field matching, long document chunking, dataset isolation, and strict original text recall rules.
How to set the configurations
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | There are many professional terms in the water utility sector, so a high matching threshold is needed to filter irrelevant content and avoid retrieving unrelated cases |
limit | Top 8–12 entries | Case review requires reference to multiple related materials, and this range can cover the main reference basis |
searchMode | "Field Matching + Full-text Retrieval" | Balances precise recall of structured project numbers and unstructured body text |
usingReRank | Enabled | After chunking long documents, reranking is required to optimize relevance sorting and improve review reference efficiency |
datasetSearchUsingExtensionQuery | Disabled | Strictly limit returning only original answers from knowledge base question-answer pairs, without triggering extended generation |
collectionIds | Specify the dataset ID corresponding to the water utility application review cases | Isolate case data from different projects, enabling retrieval scope filtering by folder |
Three common mistakes
- Phenomenon: Unable to filter knowledge base content for a specified project folder during search. Cause: The corresponding dataset was not bound via the
collectionIdsconfiguration, so the retrieval scope covered all available knowledge bases. - Phenomenon: Retrieved results do not strictly return original text answers from knowledge base question-answer pairs. Cause:
datasetSearchUsingExtensionQuerywas enabled, triggering context-based extended generation. - Phenomenon: Text chunks are lost after long document chunking. Cause: The chunk length was set beyond the average character count of individual water utility case documents, or the overlap length was insufficient, resulting in incomplete capture of cross-page content.
How to confirm the configuration is correct
- Manually select the target water utility application review case dataset, initiate a test retrieval, and confirm that only content within this dataset is returned.
- Import preset water utility approval question-answer pairs, initiate a retrieval, and confirm that returned results strictly match the original text of the preset answers.
- Upload a single long case document, check that the number of chunked text blocks matches expectations, with no content loss.
- Initiate a batch vectorization task, monitor interface call logs, and confirm that no rate limit error occurs.
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 2026-09-28.