What the data for this category looks like
Marketing content data for this category comes from gas operators’ customer service systems, pipeline network operation databases, regional policy document repositories, and past marketing activity interaction records. Data update cycles cover three tiers: real-time, daily, and weekly. Real-time sync covers behavior data such as user payments and service requests. Daily sync covers static tags such as customer gas usage tiers and regional service coverage areas. Weekly updates cover policy content such as regional gas subsidies and service upgrades. Data is divided into two categories: structured and unstructured. Structured fields include 6-digit regional codes, gas usage statistics, and marketing channel types, with units being administrative division codes, cubic meters, and text identifiers respectively. Unstructured content includes long-text materials such as promotional campaign copy and community service notices.
What constraints these characteristics impose on deployment and upgrade workflows
Based on the above data characteristics, the deployment and upgrade workflow must match multi-dimensional rule constraints. First, multi-frequency data sources require layered sync tasks. Separate sync frequencies for real-time behavior data, daily static tags, and weekly policy documents to avoid full syncs consuming excessive system resources. Second, fixed coding and unit rules for structured fields require format validation logic configured during data import preprocessing. This prevents marketing content targeting errors caused by incorrect regional codes or inconsistent gas usage units. Third, unstructured long-text materials have wide length ranges. Support adjustable segment parsing thresholds to avoid parsing failures for overly long materials. Fourth, sync of real-time user behavior data requires low-latency transmission links. This ensures marketing content can be generated quickly based on the latest user status. The upgrade phase must retain existing layered sync configurations to avoid sync logic disruption from version updates. It must also support parsing formats for newly added policy documents to ensure normal invocation of existing materials.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Gas marketing materials include long-text policy documents and complex structured tables, requiring sufficient time to complete parsing |
UPLOAD_FILE_MAX_SIZE | 1000 MB | Gas regional policy documents may include PDFs compiled from multi-page scanned documents, resulting in large individual file sizes |
maxContext | 8000–12000 characters | Gas marketing content needs to combine multiple fields such as regional codes and gas usage data; a longer context ensures content relevance |
recall count | Top 8 entries | Gas marketing requires precise matching of regions and user tiers; excessive recall leads to content redundancy |
similarity threshold | 0.75–0.85 | Low-match non-target regional marketing materials must be filtered out to avoid incorrect content delivery |
RELOAD_MODEL_ON_STARTUP | true | Offline deployment or post-upgrade requires reloading model weights to ensure normal knowledge base query functionality |
The parameter values provided on this page are common starting points for configuration. Actual values are affected by material formats, data volume, and business rules. Specific issues require case-by-case analysis, and it is recommended to test on your own samples before finalizing settings.
Three common errors
- Issue: After deployment, gas policy document images in the knowledge base fail to load normally, and the frontend page returns a 404 status code. Cause: Static resource mapping paths are not configured, and the image directory for gas materials is not mounted to FastGPT's static service directory during local deployment.
- Issue: After offline deployment and reranking model configuration, clicking model test pops up a "request error" prompt, and the curl test returns a 500 status code. Cause: The local reranking model image package is not imported into the container environment, or the environment variable
RE_RANK_MODEL_PATHdoes not correctly point to the model file path. - Issue: Knowledge base query functionality malfunctions after offline upgrade, and the large language model does not organize and return retrieved gas marketing data. Cause: The reload operation configured by
RELOAD_MODEL_ON_STARTUPis not executed after upgrade, or the vector index of the existing knowledge base is incompatible with the parsing logic of the new version.
How to confirm successful configuration
- Run the layered sync task test script to verify that the sync frequencies for real-time behavior data, daily static tags, and weekly policy documents meet the configured requirements.
- Upload a gas policy PDF containing multi-page scanned documents, check that the parsing completion time falls within the range configured by
PARSE_FILE_TIMEOUT_SECONDS, and that the parsed field formats are correct. - Trigger a knowledge base recall test, verify the matching results against the recall count and similarity threshold, and ensure only target regional marketing materials are returned.
- Restart the FastGPT service, check the console logs to confirm that model weights are loaded successfully with no error messages.
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.