What this category of data looks like
Data sources include public market operation announcements, interbank lending market transaction data, financial institution financing record filings, and third-party financial news APIs. Updates follow a daily T+1 cadence, with some real-time market fields refreshed hourly. Documents use a structured header paired with detail rows. Headers include these fields: release date, financing subject, financing amount, financing term, funding cost, and funding purpose. Units: Financing amount uses ten thousand yuan as the unit, financing term uses days, and funding cost uses annualized interest rate values.
Constraints on Knowledge Base Retrieval and Recall
High structured field proportion requires configuring differentiated recall weights for core fields like financing subject and financing amount, to avoid non-core fields interfering with matching results. Daily T+1 updates with hourly-refreshed fields require an incremental synchronization mechanism, to only update newly added or modified documents that day and avoid recalling outdated data. Fixed document structure and clear fields enable precise recall via field extraction, reducing irrelevant content mixing. Uniform units but diverse field types require verifying field unit consistency during recall, to avoid invalid cross-unit matching.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Recall count | Top 10–15 entries | Each document contains multiple groups of financing details. Too many recalled entries increase context processing overhead, while too few fail to cover complete target financing subject information. |
Similarity threshold | 0.72–0.85 | Structured field matching has high precision, so the threshold can be appropriately raised to filter non-target documents with low matching degrees. |
Chunk size | 800–1200 characters | Must cover complete single-group financing entries, avoid splitting that disrupts field relevance, ensuring semantic integrity during retrieval. |
Incremental Update Toggle | Enabled | Data updates daily with real-time refreshed fields. Incremental synchronization reduces repeated parsing overhead and ensures data timeliness. |
Field Weight Configuration | Financing subject:1.5, Financing amount:1.2, All other fields:1.0 | Core fields are high-frequency search terms for users. Raising their corresponding weights optimizes matching precision. |
Rerank result count | Top 5–8 entries | Perform secondary sorting on recall results, retain the most matching core financing entries, and simplify subsequent model processing workflows. |
The parameter values provided on this page are common starting points for configuration. Actual values are affected by material form, data volume and business rules. Each situation requires specific analysis, and it is recommended to test on your own samples before finalizing the settings.
Three Common Configuration Mistakes
- Phenomenon: A "Request failed with status code 400" error is returned when calling the knowledge base API. Cause: Uploaded financing daily report documents have non-standardized field formats, or required document metadata fields are missing from the API request body.
- Phenomenon: A large number of unstructured raw text fragments are mixed into retrieval results. Cause: Field extraction configuration is not enabled for structured documents, resulting in failure to extract core field information during parsing.
- Phenomenon: Slow retrieval response when deploying the model locally. Cause: The number of recalled entries is set too high, and the re-ranking function is not enabled, causing a large amount of redundant text to enter the model context processing flow.
How to Verify Successful Configuration
- View the knowledge base synchronization log to confirm that only newly added or modified documents on the current day are parsed, verifying that the incremental update configuration takes effect.
- Enter search terms that include core fields, check the matching degree of core fields in the retrieval results, confirming that the field weight configuration takes effect.
- Test search terms with different units, confirm that the retrieval results only include financing entries with matching units, verifying that the unit verification configuration takes effect.
- Check the retrieval response time to confirm it meets business requirements, verifying that the segment length and number of recalled entries balance overhead and effectiveness.
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.