What data for this category looks like
Glass market and yield rate related data is mainly sourced from public quotation APIs of the domestic building materials circulation association and futures exchange market data sources. There are two update schedules: spot quotations are updated once daily after market close, and futures market data is pushed hourly during trading days. The data uses structured JSON format. Each entry includes fields such as city name, product specification, benchmark quotation, latest transaction price, price change amount, and statistical date. The quotation unit is uniformly yuan per weight box. There are no extra nested levels. Field names are fixed with no multilingual variants.
What constraints these characteristics impose on the model integration and configuration workflow
The multiple update frequencies of glass market data require configuring scheduled polling rules for multiple data sources, and distinguishing pull frequencies between trading days and non-trading days. Fixed structured fields require specifying field mapping rules during model integration to avoid parameter parsing deviations. Unified quotation units require configuring unit validation logic to prevent confusion across product categories. Different data source interfaces have different authentication methods, so corresponding request headers and secret key parameters must be configured separately. The timeliness of daily report broadcasting requires configuring timeout thresholds to ensure data pulling and organization are completed within a fixed daily time window.
How to set the configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
MODEL_API_BASE_URL | http://localhost:8000/v1 | Follows the OpenAI-compatible API path format for vLLM inference services, and adapts to FastGPT's large model access standards |
DATA_FETCH_INTERVAL | 3600 seconds (futures), 86400 seconds (spot) | Matches the update frequency of corresponding data sources, avoids invalid pulls that consume system resources |
RECALL_TOP_K | Top 12 entries | Adapts to the scale of data source entries for the glass category, balances model processing load and information completeness |
PARSE_FIELD_MAPPING | {"city": "city", "spec": "spec", "price": "price", "change": "price_change"} | Matches the standard field naming of data sources, ensures the model accurately extracts target information |
REQUEST_TIMEOUT | 60 seconds | Adapts to the average response duration of glass data source interfaces, avoids pull failures caused by network fluctuations |
FASTGPT_MODEL_BASE_URL | FASTGPT_MODEL_BASE_URL=http://host.docker.internal:8000/v1 | Adapts to Docker Compose deployment scenarios, enables access to the host machine's vLLM service from within the container |
The parameter values provided on this page are common starting points for configuration. Actual values are affected by material form, data volume and business rules. Specific issues require individual analysis. It is recommended to test on your own samples before finalizing settings.
Three common configuration mistakes
- Symptom: Model call returns
400 Bad Requesterror with promptinvalid base url. Cause: When deploying via Docker Compose, the host address of the vLLM service is not correctly mapped to the container-accessiblehost.docker.internal, causing the FastGPT container to fail connecting to the inference service. - Symptom: Model call returns
404 Not Founderror, or generated content does not meet expectations. Cause: The request parameter format is not adjusted for distilled large models, and the interface call rules for full-sized large models are still used, causing the model to fail responding correctly. - Symptom: Daily report generation delay exceeds the preset threshold, triggering system timeout alerts. Cause: Pull intervals for spot and futures data sources are not distinguished, and a unified high-frequency pull rule is used, causing interface request overload and slowed responses.
How to confirm the configuration is complete
- Initiate a model call test, check if the returned results include the configured field mapping content, and verify that the field names match the data source.
- Check system logs to confirm that within the data source update period, the number of successful interface requests meets expectations, with no frequent connection failure errors.
- Compare manually pulled data source content with the fields extracted by the model, confirm that the quote unit and statistical date match.
- After adjusting the pull interval, observe system resource usage to ensure no CPU or memory overload occurs.
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.