What the Knowledge Base Data Looks Like
The knowledge base data for iron and steel smelting mainly comes from on-site process specifications, equipment operation and maintenance logs, safety operation regulations, raw material quality inspection reports, and finished product inspection documents. The data covers the entire smelting process, with differentiated update rhythms: process specifications are revised quarterly, operation and maintenance logs are synchronized daily, and safety regulations are updated annually. Most documents are a mix of structured and semi-structured content, containing fields such as process number, furnace temperature threshold, operating duration, and fault code. Common units include degrees Celsius, pascals, tons, kilowatt-hours, and similar units.
What Constraints Do These Characteristics Impose on Knowledge Base Retrieval and Recall
Differences in update rhythms across data sources require the recall link to support mixed configuration of incremental synchronization and full updates, to avoid excessive system resource usage from full refreshes. The high proportion of structured fields requires setting precise matching rules for fields such as process numbers and furnace temperature thresholds, to avoid fuzzy errors from full-text retrieval. The wide range of document lengths, from single operation and maintenance log entries to dozens of pages of process specifications, requires adapting chunking and recall logic for different lengths. Fields with multiple coexisting units require unified unit formatting before retrieval, to reduce matching deviations.
How to Configure the Settings
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Matches the precision requirements for iron and steel smelting process parameters, avoiding irrelevant operation and maintenance logs if set too low, and missing valid process suggestions if set too high |
limit | Top 6 entries | The number of results returned per recall round adapts to the information density of internal office scenarios, avoiding excessive information interfering with employee judgment |
searchMode | hybrid | Balances structured field matching and full-text semantic retrieval, adapting to knowledge bases mixing process specifications and operation and maintenance logs |
usingReRank | Enabled | Reorders recall results by semantic relevance, optimizing the display order of highly relevant content such as furnace temperatures and fault codes |
datasetSearchUsingExtensionQuery | Enabled | Expands queries for technical terms related to smelting processes, for example, associating "furnace temperature" with "furnace chamber temperature" and "molten steel temperature" |
collectionIds | Collection IDs divided by business domain | Distinguishes different knowledge base collections such as process, operation and maintenance, and safety, to accurately recall content for corresponding scenarios |
Three Common Configuration Mistakes
- Phenomenon: Calling the knowledge base API returns a reading link, and clicking the page pops up a format error prompt. Cause: The uploaded original document is in an unsupported format such as PDF or Word, or the document parsing process did not generate a normally accessible text jump link.
- Phenomenon: Knowledge base retrieval results are irrelevant to the current smelting office scenario. Cause:
collectionIdswas not configured correctly, and the recall scope covers documents from unrelated business domains, resulting in valid results being filtered out. - Phenomenon: Imported knowledge base documents have semantic breaks after chunking, and cannot fully transmit process steps. Cause: The parameter meaning of the custom chunking logic was misunderstood, and a reasonable overlap length was not set, causing key process parameters to be split.
How to Confirm the Configuration Is Correct
- Upload a typical steelmaking process specification document, retrieve "1600℃ converter temperature", and check whether the returned results include the parameter content of the corresponding process.
- View the
collectionIdsconfiguration of the knowledge base collection, confirm that only document collections for the current business domain are included, and exclude irrelevant data. - Adjust the
similaritythreshold, compare the change in retrieval result relevance before and after adjustment, and confirm that the value meets business requirements. - After enabling
usingReRank, compare the result order before and after reordering, confirm that highly relevant process parameter content is displayed first.
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.