Data Characteristics
Data for mental health products and reagents primarily comes from clinical trial reports, drug monographs, diagnostic criteria manuals (e.g., DSM-5, ICD-11), academic journal articles, and regulatory approvals and guidelines from pharmaceutical agencies. Data updates are relatively stable; new drug development or diagnostic standard revisions cause concentrated updates. Core data, such as drug mechanisms of action, indications, and contraindications, change less frequently. Document structures are typically highly standardized, presenting as structured clinical data tables, drug component lists, toxicology reports, pharmacokinetic curves, and free-text case descriptions and expert consensuses. Fields include, but are not limited to, drug name, active ingredient, dosage form, specifications, indications, dosage and administration, adverse reactions, interactions, preclinical study data, clinical study phases, efficacy indicators (e.g., HAM-D, PANSS scores), and safety indicators. Units strictly follow pharmaceutical and medical norms, such as milligrams (mg), milliliters (mL), micromoles (µmol/L), and percentages (%).
Constraints Imposed by These Characteristics on Tool Calling and Plugins
The standardized nature of mental health product data requires tool calls to precisely match fields and units. Any ambiguity or incompatibility can lead to biased results. For example, confusion over drug dosage units can directly affect recommendation accuracy. The cyclical nature of data updates means external tool connections to data sources need version management capabilities to ensure the latest, validated information is used. Free-text descriptions in clinical reports require plugins to extract key information from unstructured data, such as identifying specific keywords in patient symptom descriptions and mapping them to diagnostic criteria. Additionally, queries involving drug interactions require the toolchain to perform cross-database queries and handle the merging and conflict resolution of query results. Queries for efficacy and safety indicators require tools to parse complex statistical data and provide quantifiable results.
Configuration Guidelines
| Configuration Item | Suggested Value | Rationale |
|---|---|---|
toolCallTimeout | 60 seconds | Most external APIs respond to complex queries within 10-40 seconds. This provides a buffer against network fluctuations. |
maxToolCalls | 3 | Avoids excessively long processing chains from over-calling, improving response efficiency and reducing unnecessary external API costs. |
functionSchema | Precisely define name, description, and parameters for drug queries, component analysis, and interaction checks | Ensures FastGPT accurately understands tool functions and passes correct parameters as needed, reducing hallucinated calls. |
knowledgeBaseRecallLimit | 5 | When retrieving relevant literature or guidelines, the top 5 entries usually contain the most core and direct matching information. |
similarityThreshold | 0.85 | Ensures recalled knowledge snippets are highly relevant to mental health product queries, reducing interference from irrelevant information. |
contextWindow | 8000 tokens | Balances the model's depth of context understanding with processing efficiency, ensuring complex query background information can be included. |
Common Mistakes
- External tool return values are not obtained in the workflow: This usually happens when the
outputparameter in the tool function definition is not configured correctly, preventing FastGPT from recognizing and parsing the tool's returned data structure. - The prompt fails to effectively trigger tool calls, or calls the wrong tool: The main reason is that the
descriptionorparametersinfunctionSchemaare not clear enough or do not match the model's expectations, failing to accurately guide the model to select the correct tool. - Knowledge base retrieval results do not match the query: This often occurs when
similarityThresholdis set too low, recalling too much generalized or low-relevance content, or when the knowledge segmentation strategy fails to effectively retain key information.
Verification Steps
- For core drug query and component analysis scenarios, write test cases covering different parameter combinations. Observe whether the tools are correctly identified and called, and verify the accuracy of the returned results.
- Check FastGPT's operational logs to confirm that the parameters for each tool call request meet expectations and that the external API returns a success status code.
- Simulate user consultations multiple times to verify whether the knowledge base's recalled snippets are highly relevant to specific mental health symptoms and drug indications. Evaluate the professionalism of the answers generated by the model based on the recalled information.
The values provided are common starting points and should be measured against the reader's own samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-21.