This configuration scheme applies to FastGPT v4.17.0.
What the data for this category looks like
The data sources for this category include wood pulp and waste paper batch archives, DCS/QCS production control systems, paper machine online measurement equipment, laboratory test reports, and energy and water treatment system logs. Document types include pulp preparation sheets, beating records, paper machine operation logs, online quality curves, and sheet break and product change records. Fields include pulp ratio, beating degree (°SR), paper grammage (g/㎡), moisture content (%), paper defect type, and others. Data update frequencies are real-time (paper machine operation parameters), per shift (team production reports), per roll (single roll paper quality archive), with no fixed batch update cycle.
What Constraints These Characteristics Impose on Knowledge Base Retrieval and Recall
Multi-frequency data updates require the knowledge base to support incremental synchronization, to avoid excessive computing resource usage from full updates. High proportions of structured fields and numeric parameters require retrieval to support structured matching and numeric range queries, rather than relying only on keyword fuzzy matching. Production data for a single roll of paper is closely linked. Tender quotations need to match historical data of the same paper grade and production batch, so retrieval must support filtering by paper grade and batch fields. Multiple document types are stored together. It is necessary to distinguish between pulp preparation parameters, quality curves, business logs, and other document types, to avoid cross-interference in recall results. In high-speed continuous web production mode, parameters for a single roll of paper have strong continuity. Retrieval must avoid recalling invalid data across batches.
How to Set the Configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.80 | Matches structured numeric fields and business keywords for pulp and paper. The range is calibrated via actual testing to avoid recalling irrelevant historical pulp preparation data |
limit | Top 6 entries | Production data for pulp and paper is closely linked. Too many single recall results will exceed the model context limit. 6 entries can cover typical pulp preparation and operation parameters for a single roll of paper |
searchMode | mixedRecall | Balances numeric features of pulp preparation sheets and quality curves with business keywords in tender documents such as "120g/m² white cardboard", to improve recall accuracy |
usingReRank | Enabled | For mixed features of multiple document types, re-ranking filters out results that meet similarity standards but do not match business requirements. For example, it places valid data of the same paper grade at the top |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands keywords for tender requirements. For example, it expands "high bulk pulp preparation" into business terms corresponding to pulp ratio and beating degree, to match historical documents |
collectionIds | Bind the exclusive collection ID of the pulp and paper tender knowledge base | Avoid retrieving data from other category knowledge bases, and ensure recall results only come from this industry's business documents |
Three Common Mistakes
- Fail to retrieve corresponding content after importing external web links, with a prompt of no matching results. Web login restrictions are not handled. FastGPT cannot crawl content from authorized pages, leading to missing data in the knowledge base import.
- Receive a prompt of "no knowledge base selected" when calling the knowledge base, but normal recall works in debug preview. The knowledge base collection ID configuration in the production environment is not synchronized to API call parameters, or the environment variable is not correctly bound to the corresponding knowledge base collection.
- The number of retrieval results is far lower than expected, and contains a large number of irrelevant non-pulp and paper business data. The
collectionIdsconfiguration to limit the exclusive knowledge base collection is not set, or thesimilaritythreshold is set too low, recalling low-matching documents from other categories.
How to Confirm the Configuration Is Correct
- Upload a single paper machine operation record document, trigger retrieval, and check if the returned results contain the key fields of the document such as beating degree and paper machine speed.
- Set the
similaritythreshold to 0.75, enter the test query "150g/m² coated white board pulp preparation plan", and verify whether the paper grade matching degree of the recall results meets expectations. - After enabling
usingReRank, compare the recall result rankings before and after enabling, to confirm that high-matching business documents are placed at the top. - Check the knowledge base update log, confirm that newly uploaded pulp preparation sheets and product change records can be synchronized to the retrieval library within 10 minutes.
The values provided here are common starting points, and should be evaluated against the reader's own samples.