This document is compatible with FastGPT v4.17.0.
What data in this category looks like
Data for this category comes from PLM, ERP, MES, supplier materials, validation tests, and customer-specific requirements. Updates are triggered irregularly with each process or engineering change. Daily documents include ECN/ECR forms, change assessment reports, drawings and BOM version records, process validation reports, production switch plans, inventory disposal documents, and customer approval files. Data fields include version number, change scope, validation results, and other fields. There are no unified fixed units; fields and units vary based on the content of individual documents.
Constraints imposed on vector models and indexing
Data sources for this category span multiple systems, with heterogeneous formats and significant differences in single-document length, ranging from a few-page form to a dozens-page validation report. Chunking logic must be adapted for different structures. Update triggers from changes have no fixed cycle, and single-batch data volume fluctuates with change scale. Index refresh rhythm must support dynamic adjustment. Documents contain key fields such as version number and change scope; vector indexes must retain field-level associations to ensure retrieval accuracy. Changes are linked to multiple stages of data, so retrieval must cover cross-document associated content, and should not be limited to single-document fragment matching only.
Configuration settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Chunk length | 800–1200 characters | Adapts to the length difference of documents in this category, from a few pages to dozens of pages, to avoid chunk loss or over-fragmentation |
| Overlap length | 100–200 characters | Ensures cross-chunk change-related content is not truncated |
| Similarity threshold | 0.75–0.85 | Filters low-relevance results, and adapts to semantic matching accuracy for change-related content |
| Single search limit | Top 10 entries | Covers cross-document retrieval needs associated with changes, to avoid missing associated content |
maxToken | Set to the selected embedding model/service maximum supported input tokens | maxToken is the input-token maximum supported by the selected embedding model/service; verify it in that provider configuration. See the v4.17.2 model-field reference. It is not retrieval/LLM context or a FastGPT-wide fixed cap. |
defaultToken | 2000 token | Controls the basic length of a single embedding, balancing embedding quality and speed |
Three common configuration mistakes
- Phenomenon: Text chunk loss occurs when chunk length is set to 3000 characters. Cause: Some long documents in this category contain continuous version numbers and change-associated paragraphs. Fixed 3000-character chunking triggers the system's truncation threshold, resulting in some fragments not being correctly split.
- Phenomenon: Rate limit errors are triggered during vectorization. Cause:
defaultTokenand thread configuration are not adjusted. The single-batch embedding task volume exceeds the system's allowable rate limit. - Phenomenon: The Q&A pair extraction task remains in the "Indexing" status and cannot change to ready. Cause: The irregular change update rhythm of this category is not adapted. The index refresh trigger logic does not match the data update frequency, resulting in task backlog.
How to confirm correct configuration
- Run a chunking test on a single long document, verify that the number of split chunks matches expectations, and adjust chunk length and overlap length to meet requirements.
- Run a small-batch vectorization task for change documents, monitor embedding speed, and adjust
defaultTokenand thread configuration to meet system limits. - Initiate a retrieval test for change-related content, verify the relevance and quantity of returned results, and adjust similarity threshold and single search limit.
- Submit a simulated change index refresh task, confirm that the task status can be normally updated to completed.
The values provided are common starting points and should be measured against the reader's own samples.