What data looks like for this category
This category of data primarily originates from logging systems, Trace platforms, Metrics systems, and application SDKs. It updates once daily or per version iteration. Daily operation and maintenance documents include log collection specifications, link tracing instructions, indicator dictionaries, and dashboard configuration records. Data fields contain structured content such as timestamps, service names, TraceID, SpanID, call latency, indicator dimensions and values, alongside unstructured log text snippets.
What constraints do these characteristics impose on the knowledge base retrieval and recall link
Multi-source heterogeneous data sources require the retrieval link to support associated recall of cross-system documents, avoiding fragmented information. The high-frequency update feature requires the knowledge base to be configured with an incremental synchronization mechanism, ensuring retrieval results always align with the latest collection specifications and dashboard configurations. The presence of structured fields and professional terminology requires retrieval to balance keyword matching and semantic understanding, preventing recall failures caused by mismatched field formats. The core requirement of fault location requires recall results to include sufficient contextual evidence, allowing operation and maintenance personnel to quickly identify the root cause of issues.
How to set configurations
These configurations apply to FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Logging and observability documents contain a large number of professional terms and specially formatted fields. This threshold filters low-match irrelevant content and retains core relevant information |
limit | Top 5–15 entries | Operations duty scenarios require reference to multiple log segments and link tracing fragments. This range balances the number of recall results and context window usage |
searchMode | Hybrid retrieval | Balances keyword matching and semantic understanding, adapting to mixed retrieval needs for structured fields such as log IDs and service names, alongside general explanatory text |
usingReRank | Enabled | Semantic similarity differences between log and indicator data are relatively small. The re-ranking step can further filter recall results highly relevant to the query |
datasetSearchUsingExtensionQuery | Enabled | Operations issues often include industry abbreviations or custom terminology. Extended queries can supplement matching of relevant specification documents |
collectionIds | Bind dedicated operations and maintenance knowledge base collections | Avoid recalling non-operations documents, ensuring retrieval results focus on content related to log platforms and observability |
Three common mistakes
- When creating a knowledge base using a URL, the prompt
cannot fetch internal urlappears: FastGPT has not been configured with network permissions to access internal logging systems, or the interface corresponding to the target URL has not opened necessary access permissions. - When clicking chunk preview for a PDF file in the knowledge base, the prompt that the file content cannot be read appears: The PDF file contains encrypted content, embedded non-standard fonts, or has a corrupted file structure, causing parsing failures.
- When uploading to the knowledge base, multi-file upload is unavailable, or a network failure occurs after a single file takes too long to transcode: The concurrency threshold for batch uploads has not been configured, or the single file size exceeds the upload limit allowed by the knowledge base, causing server processing timeouts.
How to confirm the configuration is correct
- When retrieving questions related to the log platform, check whether the recall results include exclusive document content such as log collection specifications and link tracing instructions.
- Check whether the matching degree of the retrieval results matches the set
similaritythreshold, with no obvious low-match irrelevant content. - Verify that when uploading files in batch, multiple log-related documents can be uploaded simultaneously, and no timeout errors occur during single file transcoding.
- After enabling the re-ranking function, check whether the top-ranked results have a higher degree of relevance to the operations query.
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.