What this category of data looks like
Coal power integration declaration materials and file review data mainly come from coal project approval declarations, capacity verification approval forms, environmental assessment filing files, grid connection acceptance materials, and similar sources. Data update rhythm fluctuates with project milestones, with bulk updates during quarterly declaration and annual verification phases, and incremental updates for archived files on a daily basis. Documents are mostly semi-structured PDF declaration forms and structured Excel approval forms, with fixed fields including project number, coal production capacity, installed capacity, declaration date, approval status, and more. Units include tons, ten thousand kilowatts, natural days, and others.
What constraints do these characteristics impose on the knowledge base retrieval and recall link
The mixed structured and semi-structured nature of coal power integration declaration materials causes cross-row merging issues during automatic splitting. Row-level splitting rules must be configured specifically to address this. The presence of fixed fields requires retrieval to balance precise field matching and semantic recall, to avoid recalling irrelevant files. The batch update rhythm requires the index to support incremental import and batch task scheduling, to adapt to centralized data updates. The length of single files varies widely. Long-text declaration documents need to be split to fit the input limits of retrieval models, to prevent reduced recall accuracy caused by overly long single chunks.
How to set the configurations
These configurations apply to FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Coal power declaration materials have fixed approval fields. Low-match irrelevant files must be filtered to avoid interfering with the review process |
limit | Top 8–12 | Single file review needs to cover multiple related materials of the same project. Recalling too much content increases the verification burden on reviewers |
searchMode | vector+keyword | Balances precise matching of structured approval fields and semantic recall of semi-structured declaration documents, adapting to mixed-format data |
usingReRank | Enabled | Semantic similarity of long-text declaration materials may deviate. Re-ranking improves the accuracy of top results |
datasetSearchUsingExtensionQuery | Enabled | Declaration materials often contain fixed industry terms. Extended queries cover synonymous expressions and improve recall coverage |
collectionIds | Bind exclusive coal power declaration file collection | Isolates declaration materials of different projects to avoid irrelevant cross-project recall content interfering with current review tasks |
Three common mistakes
- Phenomenon:
upstream connect error or disconnect/reset before headerror occurs when uploading large-volume coal power declaration PDFs. Cause: Chunk size has not been adjusted for long-text files, and single-batch import data volume exceeds the interface processing limit. - Phenomenon: Multiple rows of content from Excel-format approval forms are merged into a single chunk after import. Cause: Custom row-separated splitting rules have not been configured. The default splitting logic does not adapt to the row-level structure of structured tables.
- Phenomenon: Retrieval speed is too slow when searching QA split indexes created via OpenAPI. Cause: The recall upper limit
limitis set too high, andusingReRankis not enabled. This leads to too many redundant results and increases retrieval load.
How to confirm the configuration is correct
- Upload a single Excel approval form for a coal power project, check if chunk results are split by row with no cross-row merged chunks.
- Initiate a retrieval request for fixed approval fields, verify that the
similarityscores of returned results fall within the 0.72–0.85 interval, filtering out low-match content. - Batch import 10 similar declaration materials, check if index completion time meets expectations with no timeout errors.
- Call the knowledge base retrieval interface, verify that returned results only include declaration files of the current project with no irrelevant cross-project content.
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.