What this category of data looks like
Shipbuilding industry party building learning Q&A data mainly comes from party building documents issued by the enterprise party committee, learning ledgers of each branch workshop party branch, annual learning files of party members, and latest instructions and requirements from superior party committees. The data update rhythm is synchronizing branch learning updates monthly, and updating superior party building policy documents quarterly. Most documents are in doc or docx format, with clear section titles and body paragraphs as their structure. Fields include the branch the file belongs to, upload time, original content, and section hierarchy information, with no additional custom unit fields.
What constraints these characteristics impose on the "knowledge base retrieval and recall" link
Reference only; verify against the target FastGPT release and deployment documentation.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Shipbuilding party building content is mostly normative text, requiring filtering of low-relevance generalized expressions to retain accurately matched knowledge points |
limit | Top 6 entries | Relevant content for a single branch party building learning question is usually concentrated in a small number of documents; excessive recall will increase context redundancy |
usingReRank | Enabled | Party building documents have clear section hierarchies, and reranking can prioritize recalling section content that best matches the question's semantics |
collectionIds | Bind the collection ID of the corresponding branch | Party building content of each branch workshop in shipbuilding is independent, limiting collections can avoid irrelevant results across branches |
datasetSearchUsingExtensionQuery | Enabled | Party building questions often contain industry-specific expressions, and optimization can improve retrieval accuracy |
searchMode | Semantic search | Party building learning Q&A requires matching semantically related knowledge points, and adopting semantic search mode first can improve matching accuracy |
Three common mistakes
- Phenomenon: The doc/docx document content returned by RAG retrieval does not correctly carry the corresponding section title, or the section title does not match the body paragraph. Cause: The recall configuration associated with the document structure is not enabled, resulting in the parsed section hierarchy information not being included in the retrieval index.
- Phenomenon: The AI-generated party building learning Q&A content exceeds the scope of the knowledge base and includes expressions not included in the knowledge base. Cause: The dataset scope limitation is not enabled, or the external generation switch is not turned off, causing the model to call non-knowledge base data sources.
- Phenomenon: The doc/docx files uploaded to the knowledge base cannot be downloaded directly, and can only be obtained by clicking "View Original Content" to jump to a fixed address. Cause: No dedicated download entry is configured, and currently only the jump link to the original content is open, with the address format being the http://fastgpt/ewchat/api/com related path.
How to confirm the configuration is correct
- Upload a doc format party building document with clear section titles, initiate a retrieval request for the corresponding section, and check whether the recall result carries the section title information.
- Limit the dataset scope of a specified branch, initiate a retrieval for a party building question across branches, and check whether the results only return document content within that branch's scope.
- Upload a test file, check whether there is a direct download entry on the interface, or whether the jump link to the original content conforms to the expected format.
- Initiate a party building question query that exceeds the knowledge base scope, and check whether the AI prompts that it cannot answer or only responds based on the knowledge base content.
The values provided here 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.