What data for this category includes
Data for this category comes from on-site first-hand resources including customer data center environments, equipment lists, software packages and network topologies. Daily outputs include deployment manuals, installation records, configuration lists, on-site check-in sheets and other documents. Data updates happen per order or phase, and align with on-site implementation progress. Data fields include equipment model, software version, IP address, configuration parameters, on-site check-in information and more. Common units include units, GB, minutes, percentage and more. Each document focuses on specific steps and results of a single on-site deployment.
Constraints on Knowledge Base Retrieval and Recall
Data for this category is mostly exclusive to a single on-site deployment. It updates dynamically with implementation progress. Knowledge base retrieval must accurately match specific equipment and environment parameters for a single deployment. This avoids irrelevant recall results. On-site deployment documents have scattered structures. They cover multiple types of fragments: installation steps, configuration details, troubleshooting records. Retrieval must support filtering results by document type and deployment phase. Data updates occur frequently. Teams must synchronize latest deployment records and configuration lists regularly. This ensures retrieval results match current on-site environments. Additionally, on-site deployment data often involves customer private environment information. Recall scope must be strictly controlled. Only assigned implementation engineers and customer IT staff can access corresponding datasets.
Configuration Settings
The following configurations are formulated based on the rules of FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | On-site deployment data mostly consists of structured parameters and step descriptions. This range filters low-similarity irrelevant recall results, and balances retrieval precision and coverage |
limit | top 8–12 results | On-site deployment documents have moderate individual content length. A small number of recall results can cover required configuration steps and troubleshooting solutions, and avoids context overload |
searchMode | vector search | On-site deployment data mostly consists of structured equipment and configuration information. Vector search can accurately match keywords and semantically related deployment content |
usingReRank | enabled | On-site deployment data contains many similar configuration parameter descriptions. Re-ranking further filters redundant results, and improves the ranking priority of key information |
datasetSearchUsingExtensionQuery | enabled | On-site deployment questions often include specific details of customer environments. Extended queries optimize the matching degree between retrieval terms and actual questions, and recall documents that better fit on-site scenarios |
collectionIds | bind the exclusive dataset ID of the current deployment project | On-site deployment data is exclusive to a single project. Binding the exclusive dataset avoids recalling irrelevant content from other projects, and ensures retrieval precision |
Three Common Mistakes
- Retrieval results do not return database or original document fragments: The cause is that
datasetSearchUsingExtensionQueryis not configured to associate custom data sources, or the result output format of the knowledge base retrieval node is not bound in the workflow. This prevents the large language model from directly referencing original document fragments. - Garbled characters appear in uploaded deployment manuals in the knowledge base: The cause is that the encoding format of the uploaded file does not match the default parsing format of FastGPT, or on-site deployment documents contain non-standard special characters that are not converted to UTF-8 encoding.
- The number of retrieval results does not match the configured
limitparameter: The cause is insufficient valid documents in the dataset, or thesimilaritythreshold is set too high. This filters most eligible recall results, leading to actual returned results being lower than expected.
How to Confirm Configurations Are Properly Set
- Upload a latest on-site deployment configuration list to the exclusive dataset, initiate a test query that includes the configuration parameters, and verify whether the retrieval results include corresponding document fragments.
- Check the startup logs of the FastGPT Admin service, confirm that
AIPROXY_API_ENDPOINT(root address of the AI Proxy service) andAIPROXY_API_TOKEN(AI Proxy administrator token) are correctly configured, and there are no startup verification failure prompts. - Adjust the
similaritythreshold and initiate multiple test queries, observe changes in the number and relevance of recall results, and confirm that the configuration takes effect. - After binding the exclusive dataset ID, initiate deployment problem queries for other projects, verify that document content from unrelated projects is not recalled, and ensure dataset isolation takes effect.
The values given 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.