What this category's data looks like
Data sources for coal chemical industry financing daily reports include publicly disclosed information from industry associations, announcements from listed entities, and filing documents from local energy regulatory authorities. The update frequency is daily. Most documents are structured tables paired with project description text, including fields such as financing party name, financing amount (unit: ten thousand yuan), financing method, fund usage, disclosure date, associated project production capacity (unit: ten thousand tons/year), environmental assessment filing number, and others. Some documents include project site photos or process flow diagram attachments.
What constraints these characteristics impose on knowledge base retrieval and recall
The daily updated data source requires the knowledge base incremental sync frequency to match the daily report release rhythm, to avoid recalling outdated content. The document structure combining multiple structured fields and unstructured descriptions requires retrieval to support both precise field matching and semantic relevance recall. The fixed unit system requires the recall logic to verify field unit consistency, to avoid invalid results caused by unit confusion. The binding relationship between associated project production capacity and financing purpose requires recall results to cover both core fields and contextual descriptions, to avoid one-sidedness from single-field matching. Some documents include filing numbers, so precise number retrieval must be supported to meet precise positioning needs in specific scenarios.
How to set configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
recall_top_k | Top 10-15 results | Coal chemical industry financing daily reports have moderate information volume per document. Too many recall results will cause redundant context, while too few will fail to cover valid information |
similarity_threshold | 0.72-0.80 | Structured fields and unstructured content are mixed. A threshold that is too low will introduce irrelevant results, while a threshold that is too high may miss valid matching items |
parse_chunk_size | 800-1200 characters | Each daily report includes multiple financing projects. The segment length adapts to the complete information block of a single project, avoiding splitting that breaks project logic |
UPLOAD_FILE_MAX_SIZE | 50 MB | Single financing daily report documents usually do not exceed 10 MB, reserving reasonable buffer space |
PARSE_FILE_TIMEOUT_SECONDS | 120 seconds | Some documents include multiple sets of project data, resulting in long parsing time. Timeouts will cause parsing failures |
field_match_weight | 0.35 | The weight of structured field matching and semantic recall must be balanced, avoiding overemphasizing field matching and ignoring contextual relevance |
The parameter values provided on this page are common recommended starting points for configuration. Actual values are affected by material form, data volume and business rules. Specific issues require specific analysis, and it is recommended to test on your own samples before finalizing settings.
Three common mistakes
- Phenomenon: When uploading a docx document containing images, the interface returns an "incorrect file format" error, or image content is not extracted after parsing. Cause: The document image parsing switch is not enabled, or the image OCR function is not enabled in the parsing configuration, causing the system to fail to recognize image content and trigger format verification exceptions.
- Phenomenon: Calling the file list interface returns an empty array, or using a curl request returns a 400 status code. Cause: The request does not carry the correct unique identifier parameter of the knowledge base, or the API permission configuration does not open the file list query permission.
- Phenomenon: After selecting the question-and-answer splitting mode, valid question-and-answer materials cannot be extracted from document images, and recall results are not associated with image content. Cause: The image OCR parsing configuration is not enabled, or the segment length is set too small, causing image content to not be included in the splitting scope.
How to confirm the configuration is correct
- Upload a single coal chemical industry financing daily report document that includes structured tables and images, check whether the parsed segmented content includes image text extraction results, and verify whether the segments completely cover the information of a single group of financing projects.
- Initiate a retrieval request, enter keywords including the financing party name and unit, verify whether the number of recall results matches the configured recall upper limit, and whether the similarity score conforms to the preset verification rules.
- Call the file list API interface, verify that the returned results include the uploaded coal chemical industry financing daily report documents, and that the number of documents matches the actual uploaded quantity.
- Configure a scheduled incremental sync task, wait for the next daily report update, and verify that the knowledge base has added the day's financing daily report documents, with no lag or omission.
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.