What this category of data looks like
This category of data primarily comes from product help center documents, customer organization structure information, training recording files, and historical ticket system Q&A lists. The update rhythm is adjusted along with customer orders or product version iterations, and is usually updated per order or per version. The documents include training courseware, official user manuals, training attendance sheets, and standardized Q&A lists. Most structures include operation step instructions, role permission divisions, and scenario-based usage cases. Fields cover training topics, adapted product versions, operation paths, Q&A scripts, and more.
What constraints these characteristics impose on the knowledge base retrieval and recall link
The data sources are scattered and include both structured and unstructured content, requiring retrieval to support mixed multi-source data recall, while also adapting to parsing logic for different formats. The update frequency is high, so an incremental recall mechanism must be configured to synchronize the latest versions of training content and Q&A information. Documents contain fields such as product versions and role permissions, so retrieval must support filtering results by version and permission scope. Some data, such as Excel-format ticket statistics, have multi-dimensional filtering requirements, so the recall logic must be able to match filtering conditions for specified fields. In addition, most training documents are step-by-step and scenario-based content, so retrieval must prioritize the relevance of operation guide-style information.
How to configure
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.80 | Matches the semantic matching accuracy required for operation steps, permission descriptions, and other content in training documents, avoiding recall of irrelevant low-relevance content |
limit | Top 8–12 entries | Matches the single-round retrieval output volume for training documents, balancing information completeness and conversation response speed |
searchMode | embedding | Prioritizes semantic retrieval mode to ensure recall relevance, given that most training documents are text-based content |
usingReRank | Enabled | Re-ranks initial recall results to optimize the display order of high-value content such as operation steps and scenario cases |
datasetSearchUsingExtensionQuery | Enabled | Generates extended queries based on user questions to cover matching scenarios for synonyms and variant expressions in training documents |
collectionIds | Bind specified collections by training topic | Isolates training knowledge bases for different customers to prevent recall of irrelevant cross-customer content |
Three common configuration mistakes
- The number of Excel data entries returned by retrieval is less than the source file. The cause is that precise filtering rules are not configured for target fields such as usage location, and only some semantically matched fragments are recalled.
- It is impossible to search for documents within a specified collection. The symptom is that retrieval returns no results after entering the collection identifier. The cause is that the
collectionIdsparameter is not configured correctly, or the collection's access permission is not open to the current session. - No corresponding results are returned when asking questions related to uploaded images. The cause is that the image text extraction function is not enabled in the knowledge base configuration, and the text content in the images is not included in the retrieval index.
How to verify correct configuration
- Upload an Excel test file containing specified fields such as usage location, initiate a retrieval request that includes the field, and check that the number of returned results matches the source file.
- Configure the specified
collectionIdsparameter, initiate a retrieval request for documents within that collection, and confirm that only content from that collection is returned. - Upload a test image containing text, initiate a retrieval request for related questions, and confirm that the returned results include the text content in the image.
- Initiate two identical retrieval requests, and confirm that there are no duplicate document fragments in the returned results.
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 2026-10-19.