Data Characteristics for This Category
High-value consumable product data originates primarily from manufacturer-provided product manuals, registration certificates, clinical trial reports, and sales channel bills of material (BOM). This data typically exists as PDF documents, Word documents, Excel spreadsheets, or structured database records. Due to their medical device classification, high-value consumables have long product iteration and regulatory approval cycles, resulting in infrequent data updates—typically quarterly or semi-annually. Major version upgrades may take longer. Document structures for product manuals usually include key information such as product model, specifications, contraindications, usage instructions, maintenance, and expiration dates. Fields are clear but may exist in multiple languages. Units involve physical measurements like millimeters (mm), grams (g), and milliliters (mL), as well as unique identifiers like batch numbers and serial numbers.
Constraints Imposed by These Characteristics on "Deployment and Upgrade"
The low update frequency of high-value consumable data means that initial knowledge base construction requires significant resources for first-time data cleaning and import, but subsequent maintenance costs are relatively low. Clear yet diverse document structures require data processing workflows to be compatible with multiple file formats and to accurately extract key information from unstructured text. For example, special handling is needed for tables and images within PDFs to ensure accurate recognition of product specifications and batch numbers. Standardization of fields and units is critical; different manufacturers may use varying expressions, necessitating unified mapping to FastGPT's internal data model to prevent incorrect consultation results due to unit inconsistencies. Given the potentially large volume of data and the presence of sensitive information, offline deployment and data security are primary considerations.
Configuration Recommendations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
UPLOAD_FILE_MAX_SIZE | 500 MB | Ensures the upload of PDF product manuals containing numerous images and charts. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Provides sufficient file parsing time for large PDF documents, preventing timeouts. |
Chunk size | 800–1200 characters | High-value consumable manuals are professional and coherent; increasing segment length retains context. |
Recall count | Top 8 entries | Ensures coverage of relevant information across multiple dimensions, such as product model, specifications, and scope of application. |
Similarity threshold | 0.75 | Guarantees the accuracy of retrieved content, preventing interference from irrelevant or low-relevance information. |
Rerank result count | Top 3 entries | Prioritizes displaying key product information most directly relevant to the user's query. |
Common Pitfalls
- Knowledge base query results lack critical parameters, such as empty product models or batch numbers. This occurs because this information might be in image format in the original document and not recognized by the text extractor.
- After deploying FastGPT, the AI agent service cannot connect to external models, reporting
get tiktoken dial tcp lookup. This typically indicates that the offline server's network configuration does not correctly resolve external AI service domain names. - After a system upgrade, some old data fails to load or displays abnormally. This happens when the new version's data model is not fully compatible with the old data format, requiring data migration or adaptation.
Verification Steps
- Upload a product manual PDF containing complex tables and multi-language content. Check if the knowledge base segmentation is complete and if key information (e.g., product name, specifications, batch number) is accurately extracted.
- Simulate a user query by asking a question that includes a product model and specific contraindications. Verify that FastGPT's response accurately mentions the corresponding product information and precautions.
- In an offline environment, access interfaces such as
http://localhost:8100/v1/chat/completionsto check if the AI agent service can correctly forward requests and receive responses, confirming correct network and proxy configurations.
The values provided are common starting points and should be measured against specific 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.