What this category's data looks like
Chemical raw material-related data primarily comes from enterprise production logs, third-party quality inspection reports, industry standard documents, and supply chain quotation sheets. Update frequency is adjusted based on business scenarios: routine quotation data updates weekly, batch quality inspection reports sync with production batches, and industry standard documents are revised every 1-2 years. A single document typically includes fields such as CAS registry number, molecular formula, density, melting point, boiling point, packaging specifications, and compliance level. Most field units are standard industrial units like g/cm³, ℃, kg/barrel, yuan/ton.
What constraints do these characteristics impose on knowledge base retrieval and recall
Chemical raw material data characteristics impose multiple constraints on retrieval and recall. There are many specialized fields and strong unique identifiers such as CAS numbers, so recall results must accurately match the target raw material to avoid confusing similar categories. Update cycles vary widely across different data types: quotation data has high timeliness requirements, so incremental training must be triggered per update cycle to prevent recalling expired quotation information. Document fields include standard industrial units, so retrieval must support unit conversion logic to match non-standard units entered by users to corresponding documents. Additionally, structural differences between batch documents and standard documents require configuring different recall weights to distinguish priority levels.
How to set configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
recall count | Top 10 entries | Chemical raw material specialized documents have many fields, so sufficient entries must be returned to cover query needs across different business scenarios |
similarity threshold | 0.82–0.88 | High precision is required for specialized term matching. A threshold that is too low may recall irrelevant similar raw materials, while a threshold that is too high may miss valid results |
chunk length | 800–1200 characters | Chemical raw material documents contain long sections of technical parameters. Too long chunks will lose local context, while too short chunks will disrupt the integrity of specialized terms |
PARSE_FILE_TIMEOUT_SECONDS | 300 seconds | Single quality inspection reports or standard documents have large content, so sufficient parsing time must be reserved to avoid timeout failures |
rerank return count | Top 5 entries | After initial recall, results must be re-sorted to focus on the top 5 most relevant entries for business invocation |
UPLOAD_FILE_MAX_SIZE | 1000 MB | Some industry standard documents have large file sizes, so the upload limit must be relaxed to adapt to batch import needs |
The parameter values provided on this page are general recommendations used as a starting point for configuration. Actual values are affected by material form, data volume, and business rules. Specific issues require case-by-case analysis, and it is recommended to test on your own samples before finalizing settings.
Three common configuration errors
- Phenomenon: The directory hierarchy of knowledge base documents returned by the API does not match the actual upload path. For example, documents stored in dir1 are displayed in the root directory. Cause: The
parentPathmetadata field was not generated correctly when uploading documents, or the configured directory parsing rules did not take effect. - Phenomenon: Knowledge base results can be recalled normally on the debug page, but some queries return no matching results via API calls. Cause: The API request does not carry the correct knowledge base binding parameters, or the
similarityThresholdvalue in the request deviates from the configuration on the debug page. - Phenomenon: Uploaded specialized chemical raw material documents cannot be recalled normally. Cause: The documents contain non-standard unit fields, and the unit-compatible parsing configuration is not enabled, so retrieval cannot match unit keywords entered by users.
How to verify correct configuration
- Upload a single chemical raw material document containing a CAS number and density parameters, check if the parsed metadata fields are complete, and confirm that the
parentPathfield matches the actual upload directory. - Initiate a test query containing specialized terms, verify that the similarity scores of the recall results fall within the preset range, and confirm that the
similarityThresholdconfiguration is active. - Trigger a batch document retraining task, check if archive records for abnormal documents are generated in the task log, and confirm that the abnormal document processing configuration is working properly.
- Call the API interface to initiate a query, verify that the directory hierarchy of the returned results matches the actual upload path, and confirm that the API parameter configuration matches the interface settings.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-14.