What this category of data looks like
Data for power marketing content comes from policy announcements, electricity package manuals, energy conservation promotion event notices, and offline service guidance documents in internal power enterprise marketing systems. Update cycles align with business adjustments: electricity price policies update irregularly per regulatory requirements, electricity packages iterate annually, and marketing events release quarterly or monthly. Document structures fall into two groups: long policy documents (including document numbers and implementation periods) and short event copy (including applicable scope and deadline). Fields include execution start date, applicable station area scope, package electricity price (yuan/kilowatt-hour), and others. Units are uniformly kilowatt-hour and yuan.
What constraints these characteristics impose on knowledge base retrieval and recall
Policy documents have strict timeliness requirements. Filter expired content using the execution start date field, and only retain currently and future effective entries during recall. Documents are bound to the applicable station area scope field. Use targeted recall by geographic dimension to avoid returning cross-region invalid content. Marketing event documents are short in length but have high release frequency. Prioritize matching content within the current event cycle during recall. Fields include numerical fields with units such as package electricity price (yuan/kilowatt-hour). Support precise matching of numerical range retrieval requirements.
How to set configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
recall count | top 8-12 entries | Power marketing content is short per piece but high in release frequency. 8-12 entries cover core information for current events, policies, and packages, avoiding redundancy |
similarity threshold | 0.72-0.85 | Power terminology has high recognizability. A threshold that is too low may return irrelevant policies, while a threshold that is too high may miss suitable marketing content. Calibrate values based on actual testing |
PARSE_FILE_TIMEOUT_SECONDS | 300 seconds | Power policy documents may include multi-page attachments, which require longer parsing time. 300 seconds covers parsing for conventional long documents |
segment length | 800-1200 characters | Power marketing content is mostly structured explanations. Overly long segments reduce retrieval accuracy, while overly short segments break context logic |
filter field matching rules | Targeted filtering by applicable station area scope and execution start date fields | Power marketing content requires matching geographic location and timeliness. Targeted filtering eliminates invalid recall results |
UPLOAD_FILE_MAX_SIZE | 1000 MB | Power policy documents may include batch attachments. 1000 MB covers conventional batch upload requirements |
The parameter values provided on this page are conventional recommendations used as a starting point for configuration. Actual values are affected by material form, data volume and business rules. Specific issues require specific analysis. It is recommended to conduct actual testing on your own samples before finalizing the values.
Three common mistakes
- Phenomenon: The reply returned by calling the chat API does not associate knowledge base documents, and the
referencefield is empty. Cause: ThekbIdsparameter is not specified in the API request to bind the corresponding power marketing knowledge base, or the knowledge base retrieval switch is not enabled. - Phenomenon: External API calls cannot access private knowledge bases, returning a 403 status code. Cause: Access permissions for the corresponding knowledge base are not enabled in the API key configuration, or the key is not bound to the correct workspace.
- Phenomenon: The knowledge base page fails to load in FastGPT 4.8.20. Cause: Local browser cache is incompatible with the new version of front-end resources, or the server-side knowledge base index has not completed initialization.
How to confirm the configuration is properly set
- Upload a test power marketing document, verify that key fields such as
applicable station area scopeandexecution start dateare correctly extracted after parsing. - Initiate a retrieval test, enter a query containing target geographic location and policy keywords, and check the timeliness and geographic matching accuracy of the recall results.
- Call the API to start a chat, confirm that the returned result includes the
referencefield and associates the uploaded test document. - Adjust configuration item parameters, compare the differences between recall results before and after adjustment, and confirm that the configuration takes effect.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-14.