What the data for this category looks like
Optical module yield rate and market trend data serving financial wealth management scenarios is sourced from publicly available optical communication industry quotation databases, securities market optical module sector market APIs, and monthly shipment data from leading manufacturers. Full basic data refresh is completed daily after market close. Real-time trading quotes are synchronized every 15 minutes. The document structure uses a structured table, including fields such as model number, manufacturer name, tax-included unit price, monthly shipment volume, and raw material cost proportion. Units are yuan per piece, ten thousand pieces, yuan per piece, ten thousand pieces, and cost proportion in decimal form, respectively.
What constraints these characteristics impose on the "tool calling and plugins" workflow
The multi-source nature of optical module data for financial wealth management scenarios requires configuring multi-data source routing rules during tool calling, to distinguish request logic between market APIs and manufacturer databases. Differences in update frequency require tool caching policies to match different expiration times for full and real-time data, to support daily report broadcast needs in financial scenarios. The diversity of structured fields requires plugin parameters to accurately match fields such as model number and unit price, to avoid null value returns that reduce broadcast accuracy. Inconsistent model naming requires plugins to include built-in standardized mapping rules, to improve call matching accuracy.
How to set configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
ragRecallTopK | Top 3 | Optical module data has many fields; excessive recall increases tool calling load, prioritize matching core information |
rerankerTopN | Top 2 | Optical module model naming is complex; reranking prioritizes returning the most matching model and unit price data |
toolCallTimeout | 60 seconds | Multi-data source calls require sufficient request time to avoid timeout interruptions |
fieldMatchThreshold | 0.85 | Similarity threshold for matching fields such as optical module model and manufacturer, reduces the probability of null value returns |
multiSourceRouterEnabled | Enabled | Optical module data comes from multiple data sources, routing to corresponding APIs is required to obtain accurate data |
ragCacheExpireTime | 86400 seconds | Full shipment data is updated daily; cache duration matches the full refresh cycle |
The parameter values provided on this page are common recommended starting points for configuration. Actual values are affected by material form, 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
- Symptom: Tool calling returns empty fields, and the API returns status code 400. Cause: The
fieldMatchThresholdparameter is not configured, or the threshold is set incorrectly, making it impossible to accurately match optical module model and manufacturer fields. - Symptom: GPU memory usage quickly grows to 6-7 GB after calling the bge-reranker model. Cause: The
rerankerTopNvalue is not limited, excessive reranking results cause the model to load additional data, and multi-process memory isolation is not enabled. - Symptom: Tool calling is triggered even when the knowledge base recalls valid content. Cause: The
toolCallTriggerThresholdparameter is not configured, or the threshold is set incorrectly, failing to distinguish the trigger logic between knowledge base recall results and tool calling.
How to confirm correct configuration
- Initiate a query containing a specific optical module model, check whether the knowledge base recall results and tool calling trigger logic conform to preset rules.
- View tool calling logs, confirm that multi-data source routing correctly matches the corresponding API, and returned fields are complete and meet expectations.
- Observe model runtime memory usage, confirm that the reranking model's memory usage is within a reasonable range.
- Trigger multiple identical queries, confirm that tool calling cache results remain stable within the corresponding cycle.
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.