What this category of data looks like
Data for this category comes from change management systems, CI pipelines, monitoring data, and business calendars. Update frequency is weekly or triggered by each change event.
Daily produced documents include change request forms, risk assessment reports, launch window plans, rollback plans, and multi-column business tables such as network segment record sheets. Document fields include change number, execution time window, risk level, rollback steps, affiliated network segment, maintainer, and other fields. Some tables have multiple associated fields. Single document length varies widely, from short requests of hundreds of characters to detailed rollback plans of thousands of characters.
Constraints imposed on knowledge base retrieval and recall
Format differences across multiple data sources require retrieval systems to support both structured tables and unstructured text. Weekly or event-driven update cycles require knowledge base synchronization mechanisms to support near-real-time updates.
Business documents with multiple columns require retrieval systems to support precise matching of specific field content, to avoid returning irrelevant entries. Mixed short and long text document structures require retrieval and recall processes to balance fast matching of short text and effective recall of long text. These processes must also address sorting optimization for long content to avoid chaotic results.
How to set configurations
The following configurations are formulated based on the parameter rules of FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75-0.85 | Change management scenarios require precise matching of business fields, filter low-match irrelevant content to ensure business relevance of retrieval results |
limit | Top 8-12 entries | Operations on-call scenarios require quick access to core information. Too many entries increase context processing load, too few fail to cover all relevant documents |
usingReRank | Enabled | Change documents contain long text segments. The re-ranking model optimizes sorting logic for long content, resolving issues where re-ranking is not triggered for long text |
datasetSearchUsingExtensionQuery | Enabled | Operations tickets often include vague descriptions. Extended queries match structured fields such as change numbers and network segments, improving retrieval precision |
searchMode | Hybrid retrieval | Supports both semantic matching and keyword matching, adapting to hybrid retrieval needs of multi-column tables and long text documents |
collectionIds | Specify the dataset ID of the corresponding change management knowledge base | Isolate knowledge base content from other business scenarios, preventing irrelevant documents from appearing in retrieval results |
Three common mistakes
- Symptom: When retrieving long text rollback plans, the re-ranking model is not triggered, resulting in chaotic result sorting. Cause: The
usingReRankconfiguration is not enabled, or the long text exceeds the input limit of the re-ranking model. - Symptom: When uploading multi-column Excel documents such as network segment record sheets, only the first two columns of content are recognized. Cause: No chunking rules are configured for multi-column tables. The default chunking logic only recognizes basic column structures.
- Symptom: Prompt template and reference template configurations are confused, leading to incorrect formatting of knowledge base reference content. Cause: The instruction logic of the prompt and the content splicing logic of the reference template are not clearly distinguished.
How to confirm configurations are correctly set
- Upload test change request forms and network segment record sheets, perform retrieval operations, check whether returned results include field content of all columns, verify that chunking and retrieval configurations for multi-column documents take effect.
- Submit a test query containing a long text rollback plan, check whether the retrieval result triggers the re-ranking model, and whether the sorting result conforms to business priority logic.
- Compare prompt template and reference template configurations, confirm that the reference template only processes content fragments returned by the knowledge base, and the prompt specifies retrieval rules and extraction requirements.
- Adjust the
similaritythreshold, submit a test query, verify whether the matching degree of retrieval results meets the precision requirements of operations scenarios.
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.