What data for this category looks like
Typical data for this category originates from central enterprise internal control systems, ERP systems, financial shared services systems, audit management systems, and procurement contract project systems. It covers documents including internal control self-assessment reports, audit notices, audit working papers, issue lists, rectification ledgers, rectification closure certificates, and more. The data updates annually. Documents contain core fields such as issue descriptions, rectification requirements, supporting materials, and closure status. Some documents include structured tables and long text passages.
What constraints these characteristics impose on the vector models and indexing link
The annual update cycle means high-frequency incremental indexing is unnecessary, so full batch indexing can be used. Documents from multiple sources and types require the indexing system to support exported content in different formats, avoiding parsing blockages. The combination of long text passages and structured tables requires chunking and indexing rules to adapt to cross-chunk associated audit issue information, preventing key rectification requirements from being split. Additionally, the core need to extract business samples and associate issues requires the index to accurately recall associated entries between vouchers and rectification ledgers, ensuring the accuracy of audit processes.
How to set configurations
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the text density of audit working papers and internal control self-assessment reports, avoiding chunk loss |
| Overlap Length | 100–200 characters | Covers cross-chunk audit issue descriptions, preventing key rectification information from being split |
| maxToken | 8192 | Adapts to the input limit of mainstream vector models, compatible with long text fields in rectification ledgers |
| defaultToken | 4096 | Used as the default context window to meet the associated retrieval needs of audit issue lists |
| Single Recall Limit | Top 10 entries | Matches the sample size range of audit sampling, avoiding interference from redundant data |
| Thread Count | 2–4 threads | Reduces the risk of exceeding embedding rate limits, adapting to bandwidth restrictions of enterprise internal networks |
Three common mistakes
- The knowledge base always displays the "Indexing" status after upload. This occurs when non-standard text formats exported from multiple systems are not adapted to, causing parsing blockages.
- A rate limit error is thrown during vectorization. This happens when the configured thread count exceeds the concurrency limit of the current environment, exceeding the call quota of the vector service.
- Text chunk loss occurs when the chunk length is set to 3000 characters. This is caused by not configuring an overlap length: long paragraphs are truncated and not correctly spliced, leading to missing key information.
How to confirm the configuration is correct
- Upload a single standard-format audit working paper, and check that the text chunks generated after indexing have no truncated or missing fragments.
- Retrieve preset audit issue keywords, and confirm that the recall results include matching rectification ledgers and associated voucher information.
- Review the vectorization logs, and ensure there are no error records of exceeding embedding rate limits.
- Trigger a full indexing task, and confirm that the task completes within a reasonable time frame with no interruptions mid-task.
The values provided are common starting points and should be measured 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 2026-09-28.