What Data for This Category Looks Like
TCM market and yield rate data is primarily sourced from daily listed trading records of domestic professional Chinese herbal medicine markets, acquisition monitoring data from authentic medicinal material producing areas, and publicly available industry market briefings. Data is updated synchronously after each trading day, and delayed to the next trading day for non-trading days. Most data documents use structured formats, including fields such as general medicinal material name, authentic producing area classification, commodity specification grade, current day's average transaction price, previous trading day's average transaction price, and number of listed trading batches on the current day. The unit for average prices is uniformly yuan per kilogram, with minor differences in pricing benchmarks across different specifications.
Constraints for Multiturn Dialogue and Prompt Engineering
TCM product categories have complex segmentation. There are cases where the same substance has multiple names, and different substances share the same name. Multiturn dialogue must first confirm the product variety, producing area, and specification step by step, to prevent the model from matching incorrect market data. Data updates follow a T+1 cycle. Prompts must clearly inform users that returned market information is from the previous trading day, and must not promise real-time data. There are many structured fields. Prompts must restrict the model to only use fields from the provided data source when generating content, and must not fabricate uncollected market information. Additionally, user queries typically require multiple rounds of parameter confirmation. Conversation context must be retained to avoid repeatedly prompting users to provide supplementary information.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | First 10 turns of dialogue or 15000 tokens | TCM product categories have complex segmentation. Retaining confirmation information about product variety, producing area, and specification from multiturn dialogue prevents repeated prompting |
topN | Top 6 entries | TCM market data has many structured fields. A small number of precise recall results can cover the core dimensions of user queries, avoiding interference from redundant information |
similarityThreshold | 0.75–0.85 | TCM varieties have cases of same substance with multiple names and different substances with same name. A reasonable threshold must be set to filter irrelevant data with low matching degrees |
PARSE_FILE_TIMEOUT_SECONDS | 300 seconds | TCM market daily report files may contain data for multiple product categories. Parsing takes a long time, so sufficient time must be reserved for structured conversion |
multiRoundMaxTurn | 12 turns | User queries for TCM market information typically require steps such as variety confirmation, specification refinement, and data comparison. 12 turns cover most scenarios |
promptTemplate | Prioritize using structured data fields provided to generate content, clearly mark the data update cycle | Restrict the model's output boundaries to avoid fabricating uncollected market information, and inform users of data timeliness |
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 specific analysis. It is recommended to test on your own samples before finalizing settings.
Three Common Misconfigurations
- Symptom: Multiturn dialogue returns market data fields that are missing or do not match the provided data. Cause: The
promptTemplatedoes not explicitly restrict the model to only use provided structured data fields, leading the model to fabricate uncollected information. - Symptom: Markdown-formatted market tables sent via WeChat channels fail to display correctly. Cause: The prompt does not specify use of the Markdown subset compatible with WeChat, or channel-specific format conversion parameters are not configured.
- Symptom: A 400 status code error is triggered when sending two messages in a single conversation during advanced orchestration. Cause: Message queue trigger conditions are not configured correctly, leading to duplicate request submissions that exceed interface limits.
How to Verify Proper Configuration
- A structured file containing market data for multiple TCM product categories may be uploaded, a multiturn dialogue initiated, and the model’s returned fields verified to exactly match the provided data.
- Queries for TCM varieties with different aliases may be submitted, and verification performed to confirm the model correctly matches corresponding market data without incorrect matches.
- A continuous 10-turn dialogue may be initiated, and verification performed to confirm conversation context is correctly retained, with no repeated prompts requesting variety or producing area information.
- A WeChat channel test may be configured, a market broadcast sent in Markdown format, and verification performed to confirm the format displays correctly.
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.