Data Overview
This category’s data primarily comes from customer interviews, bidding documents, vendor product materials, and historical project proposals. Daily processed documents fall into four categories: technical proposal books, customer requirement research forms, quotation lists, and demo scripts. These documents have distinct structural differences, including structured quotation details and lengthy technical route descriptions. Data updates occur per order or weekly. A single pre-sales project document may contain attachments in multiple formats.
Constraints for Document Parsing and Chunking
The diverse sources and varying structures of documents in this category require the document parsing process to adapt to attachments in different formats, without losing structured information. The high update frequency requires the parsing process to have fast response capabilities, to accommodate weekly or per-project iteration needs. Pre-sales documents are directly linked to bidding and project boundaries, so the chunking process must accurately retain core content such as technical routes and implementation boundaries. This prevents loss of critical information due to chunk breaks. Additionally, single documents may be large in size, so the upload and parsing process must support large files to avoid task failure due to file size limits.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunkSize | 800–1200 characters | Pre-sales technical proposals mostly consist of technical paragraphs. This length covers complete technical logic and avoids splitting across chunks |
chunkSplitter | Paragraph + semantic chunking | Documents include structured quotation lists and lengthy technical descriptions. Semantic chunking preserves context coherence |
customPdfParse | Enabled | Pre-sales documents are often complex scanned documents or multi-format collections. Custom parsing preserves formula and table structures |
PARSE_FILE_TIMEOUT_SECONDS | 300–600 seconds | Large bidding documents or multi-page proposal books take longer to parse. This prevents task interruption due to timeout |
UPLOAD_FILE_MAX_SIZE | 500 MB | A single pre-sales proposal collection may include multiple documents. This supports large file upload requirements |
chunkSettingMode | Manual configuration | Chunking needs vary significantly across different types of pre-sales documents. Manual configuration enables precise adaptation to all scenarios |
Common Misconfigurations
- Symptom: The file parsing tool fails to launch when calling the associated large language model. Cause:
customPdfParseis not enabled in the knowledge base configuration, and the parsing service is not correctly linked - Symptom: The local deployment version cannot parse complex formulas, while the online version operates normally. Cause: The formula parsing adaptation logic introduced in FastGPT v4.17.0 or later is not utilized. Older versions have insufficient chunking support for complex LaTeX formulas
- Symptom: The parsing image fails to start during docker deployment. Cause: The GPU driver directory is not correctly mounted, or the required image version parameters are not specified, leading to failure of the parsing service to run properly
How to Verify Correct Configuration
- Upload a collection document that includes a technical proposal book and a quotation list, and confirm that the parsed chunk results retain the structure and core content of all document types
- Test uploading a single document under 500 MB, and verify that no timeout or file size limit errors occur during upload and parsing
- Call the associated knowledge base in a conversation, and verify that technical route and implementation boundary descriptions in the proposal can be accurately extracted
- Check the background task logs, and confirm that the parsing task duration does not exceed the configured
PARSE_FILE_TIMEOUT_SECONDSthreshold
The values provided here are common starting points. They should be measured against relevant sample datasets.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-10-19.