What the data for this category looks like
Data sources include inventory and sales ledger systems of steel traders, credit approval ledgers of partner financial institutions, and daily public quotation data from spot markets. Updates run daily, with full business data for the previous calendar day. Two types of documents are included:
- Structured Excel daily reports, with fields including trader code, steel category, daily outbound volume, pledged inventory value, financing credit balance, daily repayment amount, and more. Units are no identifier for codes, text for categories, tons for volume, and Chinese yuan for amounts.
- OCR-recognized text from scanned paper pledge documents, which have no fixed format and scattered fields.
What constraints these characteristics impose on the knowledge base retrieval and recall link
Structured daily reports have clear fields and fixed units. Retrieval must support field-level exact matching to avoid semantic recall mixing indicators with different units. Unstructured document text from OCR has recognition errors and scattered fields. Entity extraction and field mapping must be completed before the text can be included in the recall scope. Daily full data scale grows with the number of traders. The recall link must support fast filtering by time range and trader code, to avoid returning irrelevant historical data. Core indicators of financing daily reports are closely linked. Multi-field combined retrieval must be supported, such as matching both steel category and financing balance at the same time.
How to set the configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
PARSE_FILE_TIMEOUT_SECONDS | 300 seconds | Paper documents processed via OCR typically require longer parsing times, to avoid task interruption from early timeout |
UPLOAD_FILE_MAX_SIZE | 200 MB | Total size of a single structured daily report plus associated OCR attachments usually does not exceed this value, to fit most business scenarios |
maxContext | 800–1200 characters | Core indicator text for financing daily reports falls within this range. Longer text introduces irrelevant redundant information, while shorter text loses key content |
Recall count | Top 6 entries | A standard single financing daily report has approximately 5-7 valid indicator entries. Too many entries increase context processing overhead |
Similarity threshold | 0.72–0.80 | Balances OCR recognition errors and semantic matching accuracy. A threshold that is too low introduces irrelevant results, while a threshold that is too high misses valid matches |
SYNC_INTERVAL_HOURS | 24 hours | Matches the daily update frequency of financing daily reports, to keep knowledge base data synchronized with business data |
The parameter values provided on this page are common starting points for configuration. Actual values are affected by document format, data volume, and business rules. Specific issues require targeted analysis. Testing against local samples is recommended before finalizing values.
Three common mistakes
- After selecting the target knowledge base during retrieval, no optional values appear in the reference variable dropdown menu. The cause is incomplete field mapping configuration for structured daily reports, where column names from Excel tables are not extracted as callable variable fields.
- An API call to insert documents into the knowledge base returns a 413 Request Entity Too Large status code. The cause is that the total size of the uploaded documents exceeds the value set for
UPLOAD_FILE_MAX_SIZE, exceeding the single-file limit allowed by the system. - Recall results include historical financing daily report data older than three days. The cause is that no filtering rule for data update time is configured, and synchronization does not retain only valid data updated on the current day.
How to confirm configurations are correct
- Upload a standard steel trade financing daily report Excel file. Check that the extracted fields after parsing exactly match the table column names, to confirm field mapping configuration is active.
- Submit a retrieval request that includes a specified trader code and steel category. Check that returned results only include entries matching the target conditions, to confirm filtering rules are configured correctly.
- Wait for one full synchronization cycle. Check that the update time of documents in the knowledge base matches the actual update time of business data, to confirm synchronization cycle configuration is active.
- Manually upload a piece of OCR-recognized document text. Submit a retrieval request for the corresponding indicator. Check that relevant content is matched correctly, to confirm entity extraction configuration is active.
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.