What data for this category looks like
Data related to business handling in the basic chemical industry mainly comes from internal enterprise production process specifications, equipment operation and maintenance manuals, administrative approval process specifications, and compliance inspection records. The data update rhythm adjusts with process iteration and compliance requirements. Regular process documents are reviewed and updated quarterly. Temporary approval processes are updated immediately with project changes. Most documents are a mix of structured and semi-structured content. They include fixed fields such as process parameters with units like MPa, ℃, tons per hour, approval node sequence, equipment models, and compliance standard numbers. Some documents also include on-site equipment photos and process flow diagrams.
What constraints these characteristics impose on knowledge base retrieval and recall
Structured parameter and unit fields in basic chemical industry data require retrieval to support both semantic matching and precise field matching. This avoids recall errors caused by unit conversion discrepancies. Flow diagrams and on-site equipment photos attached to documents need to support hybrid text-image retrieval logic. This prevents missing visual process guidance when only text is retrieved. When chunking long-text process specifications, retain contextual association between adjacent nodes. This prevents loss of complete approval chain information after chunking. The version field of compliance documents must be included in retrieval filter conditions. This avoids returning outdated old version process specifications.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.80 | Matches the semantic matching accuracy required for parameter-based text in the basic chemical industry. Avoids recalling irrelevant processes if set too low, and misses similar process steps if set too high |
limit | Top 8 results | Nodes in basic chemical industry business handling processes typically do not exceed 6. Retaining 8 results can cover complete processes and alternative solutions |
searchMode | hybrid | Supports semantic retrieval and precise field matching. Adapts to hybrid retrieval needs for structured parameters and semi-structured process text |
usingReRank | Enabled | Basic chemical industry documents contain a large number of parameter details. Re-ranking can filter results that are semantically similar but do not match parameters |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands synonyms for business handling-related questions. For example, expanding "hot work approval" to "hot work permit handling process" |
collectionIds | Bind exclusive knowledge base ID for basic chemical industry | Isolates knowledge base data from other categories. Avoids recalling documents from non-business domains |
Three common mistakes
- Embedded flow diagrams in retrieval results fail to load normally. The cause is that image resources were not uploaded to the bound static resource directory when uploading documents, or cross-domain permissions for image access were not configured.
- Semantic retrieval scores are abnormally high when using the
bge-m3vector model. The cause is thatusingReRankwas not enabled, and a reasonablesimilaritythreshold was not set. This results in low-relevance long-text parameters being misjudged as high-match results. - The order of business nodes returned by retrieval is chaotic. The cause is that contextual association between process nodes was not retained during chunking. The approval chain order of the original document was lost after splitting the text chunks.
How to confirm correct configuration
- Use the FastGPT v4.17.0 platform, upload a test document containing process parameters and diagrams, initiate retrieval for questions related to "hot work approval process", and check whether embedded images in the results load normally.
- Adjust the value range of
similarity, compare retrieval results under different thresholds, and confirm that matching accuracy meets the screening needs of the business scenario. - After enabling
usingReRank, verify whether the re-ranked results prioritize process documents with higher parameter matching degrees. - Check the source identifier of retrieval results, confirm that only documents from the bound exclusive basic chemical industry knowledge base are included, and no data from other business domains is mixed in.
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.