Data Characteristics for this Category
Supplier audit quality documents originate from various sources: supplier-provided quality system files, production records, inspection reports, non-conformance records, and on-site audit reports. These documents typically exist in multiple formats, such as PDF, Word, and Excel, with some potentially being scanned images. Data update frequencies vary; supplier qualification documents might update annually, while production batch records generate in real-time with production cycles. Document structures differ: quality manuals are hierarchical, while batch records are usually flat tabular data. Fields include batch number, production date, expiration date, inspection items, inspection results, judgment criteria, and deviation descriptions. Units involve mass (kg, g), volume (L, ml), concentration (%), and time (h, min). Some fields may contain free-text descriptions.
Constraints on "Reference and Traceability" from these Characteristics
The multi-source and heterogeneous nature of supplier audit documents requires robust multi-format document parsing capabilities from the knowledge base, with high accuracy for text recognition in scanned images. Inconsistent update frequencies necessitate careful attention to document version and timeliness when citing, ensuring references point to the latest or specific historical records. Documents contain a mix of structured and unstructured data, with significant variability in field and unit standardization. This challenges text segmentation and semantic understanding, requiring the model to accurately identify key information and extract data from complex tables. The reference and traceability function must precisely point to specific page numbers or paragraphs in original documents to support audit traceability requirements, avoiding vague references that only provide document names without specific content locations.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk size (Segment Length) | 500–800 characters | Balances document context and retrieval efficiency, preventing key information dilution in overly long segments. |
Chunk Overlap Length (Segment Overlap Length) | 100 characters | Ensures continuity of information across segments, capturing complete semantics that might be split at boundaries. |
Similarity threshold (Similarity Threshold) | 0.75 | Filters out low-relevance document fragments, improving recall accuracy and reducing noise. |
Recall count (Recall Count) | 8–12 entries | Provides sufficient contextual information for the large language model while controlling inference costs. |
maxContext | 3000–4000 tokens | Accommodates the complexity and detail requirements of audit documents, offering an ample context window. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Handles large PDFs or documents with complex tables, preventing parsing timeout failures. |
Three Common Mistakes
- Symptom: The model's response does not cite web search results, but the web search function tests correctly. Reason: The system is configured for non-tool-calling mode. The large language model defaults to retrieving only from the local knowledge base and is not authorized to use the web search tool.
- Symptom: Knowledge base document entries cited by the model are only listed at the bottom of the response, without specific paragraph or page number localization. Reason: The knowledge base did not retain precise location metadata during vectorization, or the frontend display layer does not implement deep linking functionality.
- Symptom: When calling the FastGPT chat interface, the response lacks the cited knowledge base ID information. Reason: API call parameters did not explicitly request return of citation details, or the API interface design does not include this information in the standard response body.
How to Confirm Correct Configuration
- Upload a supplier audit report containing complex tables and multiple pages. Ask questions about key data within it. Verify that the document citations in the model's response precisely point to specific page numbers and paragraphs within the report.
- Simulate an audit scenario by asking about the detailed processing flow for a non-conforming batch. Check if the model can synthesize information from different document types (e.g., non-conformance records and batch production records) and provide multiple citation sources.
- Make minor updates to a key document in the knowledge base. Then, ask questions about the relevant content. Verify that the model cites the latest version of the document and can distinguish between old and new versions.
Note: The values provided are common starting points and should be measured against specific samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-21.