What the data for this category looks like
The data for this category typically comes from national public resource trading platforms, natural resource trading systems, property right exchange systems, and state-owned asset supervision systems, with a daily update frequency. Commonly processed documents include transfer announcements, trading notices, bid application documents, margin vouchers, transaction confirmation letters, listing result announcements, property right transfer specifications, and more. Most documents contain fixed fields such as subject ownership information, trading subject location, margin amount, bid qualification requirements, and others. Some Excel data includes over 10,000 rows of detailed entries, and field units include square meters, ten thousand yuan, parcels, and similar units.
What constraints do these characteristics impose on the document parsing and chunking step
Batch documents updated daily require the parsing process to be efficient. A single document may contain up to 100,000 Chinese characters or 15,000 rows of data, which imposes requirements on chunk granularity and parsing time. Key content such as ownership information and bidding rules in documents must be split accurately to avoid losing associated logic across chunks. Excel-format trading detail data has dense fields. Default chunking parameters may cause a single chunk to contain too much irrelevant content, which affects the accuracy of subsequent RAG matching.
How to set the configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Adapts to the field density of property rights trading documents, avoids matching deviations caused by including too much ownership information in a single chunk |
chunkSplitter | Split by paragraph + semantics | Matches the structured paragraphs of documents such as transfer announcements and trading notices, retains key information such as bidding conditions and margin rules |
PARSE_FILE_TIMEOUT_SECONDS | 300 seconds | Addresses parsing time for single documents containing 100,000 Chinese characters (Word) or 15,000 rows of data (Excel) |
customPdfParse | Enabled | Accurately extracts subject information from property right transfer specifications, avoids field loss caused by format confusion |
UPLOAD_FILE_MAX_SIZE | 1000 MB | Adapts to the demand for batch uploading trading documents |
chunkSettingMode | Manual configuration | Adjusts chunking rules separately for Excel documents, avoids oversized data blocks caused by default parameters |
Three common configuration mistakes
- Incomplete command-related prompt text is returned when parsing Excel documents. This occurs because default chunking parameters do not adapt to data with over 10,000 rows, and a single chunk contains too many headers and row data, causing the large language model to fail to identify valid requests.
- Document parsing tasks time out. This occurs because the
PARSE_FILE_TIMEOUT_SECONDSparameter is not adjusted, and parsing time for a single large document exceeds the default threshold. - RAG knowledge base classification matching accuracy is insufficient. This occurs because chunk granularity is too large, merging paragraphs with different bidding conditions, leading to semantic confusion.
How to confirm the configuration is correct
- Upload a single Excel document with 15,000 rows, check the parsing task status in FastGPT v4.17.0, and confirm there are no timeout errors.
- Randomly extract fragments from property right transfer specifications, verify that the chunking results retain independent information blocks such as subject ownership and valuation requirements.
- Test the RAG classification matching task, enter a query related to bidding qualifications, and confirm that the returned chunk content matches the query semantics.
- Check the uploaded file size limit, confirm that it supports single uploads of trading documents matching the business scale.
The values listed here are common starting points and should be evaluated against the user’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.