Data Characteristics in this Category
Clinical Decision Support (CDS) systems primarily use data from authoritative medical guidelines, treatment protocols, drug inserts, medical literature databases, and hospital-specific SOP documents. This data typically combines structured formats (e.g., clinical pathways, drug dosage tables) and unstructured formats (e.g., PDF guidelines, text-based SOPs). Update frequency varies: guidelines and drug inserts may update annually or more often, while medical literature is continuously published. SOP document update cycles depend on internal hospital management, usually quarterly or semi-annually. Documents have complex structures, containing numerous specialized terms, abbreviations, charts, and cross-references. Fields and units are highly specific; for example, drug dosages involve milligrams (mg), milliliters (mL), units (U), and lab results involve moles (mol), liters (L), international units (IU), often with specific reference ranges.
Constraints Imposed by these Characteristics on "Deployment and Upgrades"
The complexity of CDS data places high demands on knowledge base construction. Large volumes of unstructured text require efficient text extraction and chunking strategies to ensure accurate RAG retrieval. The presence of specialized terminology and abbreviations necessitates strong semantic understanding from the model, requiring consideration of the pre-trained model's adaptability to the medical domain during deployment. The periodic nature of data updates means the knowledge base must support incremental updates and version management, avoiding full rebuilds with every update. The specificity of fields and units dictates that text embedding model configuration must optimize for numerical and unit recognition to ensure answer accuracy. For instance, when a user asks about "penicillin dosage," the system should correctly identify and match units like mg/kg. Furthermore, the challenge of integrating multi-source heterogeneous data increases the complexity of the data preprocessing phase, requiring more resources for data cleaning and standardization.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 32000 token | Medical documents are dense; a longer context helps capture complex logic and multiple references. |
Chunk size (Chunk Length) | 800–1200 characters (characters) | Balances semantic integrity and retrieval efficiency, preventing context loss from over-chunking or noise from overly long chunks. |
Similarity threshold (Similarity Threshold) | 0.78 | Clinical decisions require high accuracy; a high threshold reduces the recall of irrelevant or low-confidence information. |
Rerank result count (Rerank Return Count) | Top 5 entries (top 5) | After reranking, the top few results typically contain the most relevant key information, reducing the model's processing burden. |
UPLOAD_FILE_MAX_SIZE | 500 MB | Accommodates PDF guideline documents with numerous charts or scanned images, ensuring large files can be uploaded and processed. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds (seconds) | Complex PDFs or scanned documents require significant time for OCR and text extraction; this provides ample time to prevent parsing interruptions. |
Three Common Pitfalls
- Symptom: In knowledge base query results, drug dosages or lab indicators show mismatched values and units or hallucinations. Reason: The text embedding model failed to fully understand the association between medical values and units during training or inference, or the chunking strategy separated values from their units.
- Symptom: After deploying a local model, FastGPT cannot access the model service via a specific IP address, or tool calling functionality fails. Reason: The local model service (e.g., Ollama) is configured to bind to
127.0.0.1, preventing external or containerized FastGPT from connecting; or the model itself lacks tool calling capabilities. - Symptom: Uploaded medical guideline PDF files fail to parse, or parsed content is missing numerous chart descriptions. Reason: The PDF parser has insufficient capability to extract text from complex layouts, scanned documents, or embedded images, failing to correctly identify and extract all valid information.
How to Confirm Correct Configuration
- Upload multiple medical guideline PDFs containing complex charts and specialized terminology. Check if the parsed text is complete and logically coherent, paying special attention to whether values and units are correctly extracted.
- Conduct multi-round question-and-answer tests for typical clinical questions (e.g., "contraindications for a certain drug," "diagnostic criteria for a specific disease"). Evaluate the accuracy, completeness, and whether the answers cite correct knowledge sources.
- Simulate the data update process, for example, by uploading a new version of an SOP document. Verify that the knowledge base can identify and prioritize the latest version of information for answering queries.
- Check network connectivity between FastGPT and locally deployed LLM services. Ensure that model API response times are within expectations and that tool calling functions (e.g., drug lookup) can be triggered and return results correctly.
Note: The values provided are common starting points. Measure them against your 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.