What the data for this use case looks like
Document format proofreading data comes from official original documents and pending proofreading and circulating drafts from Party and government organs, state-owned enterprises and public institutions. Updates are triggered on demand based on the document production cycle, with no fixed scheduled update frequency. Single document character counts range from hundreds to thousands. Structures strictly comply with the Chinese national standard GB/T 9704-2012 for official document formatting, and include fixed fields such as document number, title, main recipient, text hierarchy, attachments, and signature block. Format requirements are clearly defined, such as title using 2nd-size small standard song font and body text using 3rd-size imitation song font. There are no redundant custom format items.
What constraints do these characteristics impose on knowledge base retrieval and recall
Fixed official document structure requires retrieval to match both content fields and format rule fields, to avoid retrieving irrelevant format examples. On-demand pending proofreading drafts mean the knowledge base must support recall capability for temporarily importing a single reference document, without requiring bulk import of all data in advance. Wide variation in single document character counts requires retrieval chunking strategies to adapt to format fragments of different lengths, avoiding breakage of complete format rule units. Strict format verification requirements mean recalled results must accurately correspond to specific format clauses, without generic general advice.
How to configure settings
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the precision requirement for official document formatting, avoids retrieving irrelevant format examples |
limit | Top 3–5 results | Document proofreading only requires a small number of typical compliant examples; too many results will interfere with judgment |
searchMode | Exact match | Adapts to the field matching requirement of fixed official document structure, avoids fuzzy retrieval |
usingReRank | Enabled | Performs secondary ranking on retrieved format examples to improve relevance of compliant clauses |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands retrieval keywords based on format fields of the pending proofreading document, covering more compliance scenarios |
collectionIds | Specified by document type | Only retrieves format knowledge base collections corresponding to the target document type (such as notice, report), narrowing the retrieval scope |
Three common configuration mistakes
- Symptom: GPT-4o-mini call error logs appear during workflow runtime, even though the model call was not actively configured. Cause: The
limitparameter was not set correctly; too many recalled results automatically triggered an unconfigured model completion call. - Symptom: Custom table of contents in imported official documents are automatically deleted. Cause: The automatic cleaning rule for knowledge base documents was not disabled; non-standard format table of contents nodes were cleaned by default.
- Symptom: Knowledge base question-answer pair extraction tasks remain stuck in training, with no corresponding records in call logs. Cause:
searchModewas not correctly configured as exact match, causing chunked content to fail to generate training samples properly.
How to confirm correct configuration
- Upload a pending proofreading official document template, run retrieval and recall, and verify that returned results include format compliant examples for the corresponding document type.
- View the retrieval configuration panel, confirm that values for configuration items such as
similarityandlimitmatch the preset scheme. - Trigger a question-answer pair extraction task, check that the task completes within a reasonable time frame with no abnormal stalling.
- Test importing an official document containing a custom table of contents, confirm that the table of contents nodes are not automatically deleted.
The values listed here are common starting points, and should be evaluated against the user’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.