What this category of data looks like
Sources include real-time process parameters collected from production sites, inspection records generated during quality inspection, and material consumption and production ledgers from enterprise ERP systems. Update cadence: real-time process data refreshes every 15 minutes, daily quality inspection ledgers are finalized by 24:00 on the same day, and monthly summary ledgers are generated in bulk at the end of each month. Each ledger entry includes fields such as furnace number identifier, smelting temperature value, material input quantity, production batch number, and energy consumption value. The corresponding units for these fields are degrees Celsius, kilograms, batches, and kilowatt-hours, respectively. Data is stored in structured table format, with some associated process guidance documents provided as supplementary explanations.
What constraints these characteristics impose on the knowledge base retrieval and recall link
Real-time process data has a high update frequency, which requires the retrieval link to support incrementally synced dataset configurations to avoid recalling expired data. There are many structured fields with clear units, which requires associating field dimensions during retrieval to accurately recall results and prevent invalid data across furnace numbers or energy consumption units from being included. Monthly summary ledgers and real-time process data are datasets with different time granularities, so separate independent knowledge base collections must be established to prevent mixing data from different cycles. Supplementary process guidance documents are long-form content, so retrieval must adapt to long-text chunking and recall rules to ensure core process parameters are accurately matched.
How to set the configurations
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Non-ferrous metal smelting ledger data has high field correlation. A setting that is too low will recall irrelevant furnace number data, while a setting that is too high will miss matching process parameter records |
limit | Top 8–12 entries | Each ledger entry has many fields. Retrieving too many entries will exceed the context window limit, while retrieving too few will fail to cover complete process-related information |
searchMode | Hybrid vector retrieval + keyword retrieval | Both keyword matching for structured fields and semantic recall for associated process documents are required. Hybrid mode adapts to the two-dimensional retrieval requirement |
usingReRank | Enabled | There are many similar records with the same furnace number and parameters in ledger data. Reranking can filter redundant results and improve the recall priority of core data |
datasetSearchUsingExtensionQuery | Disabled | Ledger data has structured, fixed fields. Extended queries will introduce irrelevant process description content and interfere with accurate retrieval |
collectionIds | 3 collection IDs split by data cycle | Real-time process, daily quality inspection, and monthly summary data require independent recall to avoid mixing cross-cycle data |
Three common mistakes
- Phenomenon: Retrieval results include unfiltered redundant data across the entire table, with fields beyond the preset ledger scope. Cause: The complete data table was directly imported as the knowledge base source, and no specific columns to index were specified, resulting in irrelevant fields being included in the retrieval scope.
- Phenomenon: The generated report includes process parameters not included in the knowledge base, which do not match the actual ledgers. Cause: No reasonable
similaritythreshold was set, and non-target data with insufficient matching scores was recalled, causing the generated content to deviate from the original ledgers. - Phenomenon: Furnace number queries across sessions cannot be continued; furnace number information must be re-entered for each retrieval. Cause: No retrieval association configuration for session context was enabled, resulting in each retrieval being executed independently and unable to reuse historical query conditions.
How to confirm the configuration is correct
- View the index field list of the knowledge base dataset, confirm that only the preset ledger fields are included, with no redundant column data.
- Perform a single furnace number retrieval test, verify that the
similarityscores of returned results all fall within the configured range, and that no content beyond the knowledge base scope is included. - Initiate two consecutive queries for the same furnace number, confirm that the second retrieval automatically associates the furnace number condition from the first query, without requiring re-entry.
- Check the number of returned retrieval results, confirm that it matches the interval specified in the
limitparameter configuration.
The values provided here are common starting points and should be evaluated 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.