What Data for This Category Looks Like
Core data for this category originates from ERP systems, supplier shipping documents, logistics tracking information, and equipment manufacturer backends. Daily circulating documents include purchase orders, delivery acceptance forms, serial number lists, and invoice materials. Fields cover purchase order numbers, equipment models, serial numbers, delivery dates, acceptance conclusions, and additional relevant fields. Data updates use single-batch purchases as the minimum cycle. Delivery data under a single project is synchronized once weekly. Single-file sizes typically fall within standard business ranges.
Constraints for Vector Models and Indexes
Data for this category spans multiple sources with notable field differences. Structured fields such as purchase order numbers and serial numbers are mixed with unstructured text such as acceptance notes. Index configurations must be adapted to support mixed data formats. Data updates follow a single-batch purchase minimum cycle. Incremental index trigger logic must align with delivery nodes to avoid excessive system resource usage from full reindexing. Single batches contain a large number of files, but individual file sizes are manageable. Set a per-indexing file limit to prevent processing timeouts. Retrieval must focus on core business fields including equipment models and serial numbers to prevent irrelevant content from interfering with matching results.
How to Configure Settings
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the standard text length of purchase orders and delivery acceptance forms, avoids overly fragmented single chunks |
| Overlap Length | 100–200 characters | Retains key information from adjacent chunks, prevents core fields such as serial numbers and equipment models from being split |
| Similarity Threshold | 0.75–0.85 | Meets precise matching requirements for equipment models and serial numbers, aligns with vector similarity distributions of mainstream Chinese embedding models such as bge-large-zh-1.5, reduces mismatch probability |
maxToken | 8192 token | Supports context integration for long documents, meets text length requirements for invoice materials and logistics notes |
| Single Search Limit | Top 8 results | Controls the number of returned results, avoids excessive retrieval results impacting acceptance verification efficiency |
defaultToken | 256 token | Sets the default embedding length for short single-segment text, adapts to short field content in serial number lists |
Three Common Configuration Errors
- Phenomenon: Index tasks remain in the "indexing" state indefinitely, with no progress updates or error messages. Cause: No reasonable file count limit is configured for single-batch indexing. Prolonged chunk processing time for multi-source mixed data causes processing blocking.
- Phenomenon: Retrieval results return content from two knowledge bases at once, without prioritizing the specified knowledge base per preset priority rules. Cause: Knowledge base priority binding is not enabled in retrieval configurations, or priority rules are not linked to core business fields such as equipment models and purchase order numbers.
- Phenomenon: Knowledge base index switching returns a timeout error, indicating processing duration exceeds the allowed threshold. Cause: Total file length for a single indexing operation exceeds the configured limit, or
maxTokenvalue is too large. This leads to extended embedding processing time, triggering FastGPT timeout restrictions.
How to Confirm Proper Configuration
- Upload a single purchase order document. Access the index management page in FastGPT v4.17.0 to review chunking results, confirm chunk length and overlap length match preset configurations.
- Submit a retrieval request containing equipment model keywords. Verify that returned result similarity scores fall within the preset reasonable range.
- Create two test knowledge bases, upload corresponding business documents, then submit a retrieval request to confirm results prioritize the specified high-priority knowledge base.
- Batch upload 3 to 5 business documents of different types, run an indexing operation, confirm the processing flow does not experience progress-free blocking.
The values provided are common starting points, and should be evaluated 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 2026-10-19.