Deployment and Upgrade for Traditional Chinese Medicine

What this category's data looks like

TCM yield rate and market trend data comes primarily from public datasets published by the National Traditional Chinese Medicine Circulation Traceability Platform, and official market interfaces from local professional TCM markets. Updates run after daily trading sessions close, to publish the daily settlement price and price change for individual products. Weekly, aggregated yield data for industry sectors is generated. Each data entry includes fields such as common product name, statistical date, daily settlement price, daily price change, daily trading volume, and affiliated TCM category. Field units are as follows: product name is a string, statistical date uses date format, settlement price is in yuan per kilogram, price change is in yuan per kilogram, trading volume is in kilograms, and category is a string.

Constraints for deployment and upgrade

Requirements for connecting multiple data sources mean routing rules must be configured during deployment, to distinguish market data from different sources. The daily update frequency means scheduled sync task intervals must be adjusted to match the trading day update schedule. The multi-field document structure means precise field extraction rules must be configured during knowledge base upload, to avoid redundant or missing data. The large number of products and growing data volume as categories expand mean vector database sharding and index configurations must be adjusted during deployment, to prevent reduced retrieval performance. During upgrades, new TCM sub-category data fields must be supported, to avoid conflicts between parsing logic and existing configurations. Context window adjustments from large model version updates must also be accommodated.

Configuration recommendations

Configuration Item	Recommended Value	Rationale
`PARSE_FILE_TIMEOUT_SECONDS`	`600 seconds`	TCM market data often consists of batch multi-product documents, which take longer to parse. Extend the timeout to prevent task interruptions
`maxContext`	`8000–12000 characters`	Multiple product historical market data must be included as context, to meet large model inference requirements
`RECALL_TOP_N`	`Top 10–15 entries`	There are many TCM products. Sufficient relevant data must be recalled, while avoiding context overload
`UPLOAD_FILE_MAX_SIZE`	`500 MB`	Bulk TCM market data CSV files are often large. Increase the upload limit
`SCHEDULE_INTERVAL`	`86400 seconds`	Market data updates once daily, matching the sync task schedule frequency
`VECTOR_DB_SHARD_SIZE`	`100000 entries`	There are many TCM products. Sharded storage improves vector retrieval efficiency

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 individual analysis. It is recommended to test on your own samples before finalizing settings.

Three common mistakes

Scenario: After deploying a 70B parameter model, the application responds with a timeout and returns a 504 Gateway Timeout status code. Cause: No reasonable value was configured for the maxContext parameter for the large model, leading to excessive context that triggers model inference timeout.
Scenario: After uploading TCM market data, some fields appear as null in the knowledge base. Cause: No extraction rule for the corresponding field was configured in PARSE_RULE, so the parsing engine cannot recognize preset fields such as closing_price.
Scenario: When using OneAPI to connect a locally deployed large model, the application debug returns an LLM connection failed error. Cause: The API_BASE_URL parameter was not correctly filled in FastGPT's LLM configuration, pointing to the actual deployment address of the local model.

How to confirm correct configuration

Upload a single TCM market data file, check that all configured fields are fully extracted in the parsing results, and confirm the parsing task does not trigger a timeout.
Initiate one application call, check that the response time meets expectations, and adjust the maxContext and PARSE_FILE_TIMEOUT_SECONDS parameters to reasonable ranges.
Configure a scheduled sync task, manually trigger one sync, and check that the data source successfully pulls the latest market data.
View the vector database index statistics, confirm that the number of recalled entries matches the RECALL_TOP_N parameter configuration.

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.