Data Characteristics
Psychiatric product data comes from diverse sources. These include clinical trial reports, drug inserts, medical guidelines, patient education materials, and academic papers. Document update frequencies vary. Drug inserts and medical guidelines update every few months to several years, driven by regulatory requirements or clinical data. Clinical trial reports release primarily before drug market entry, with subsequent updates for supplementary research.
Document structures typically include standard sections like indications, dosage, contraindications, adverse reactions, pharmacology, toxicology, and clinical study data. Fields extend beyond drug names, active ingredients, and manufacturers. They also cover disease diagnostic criteria (e.g., ICD-10 codes), symptom descriptions, scale scores (e.g., HAM-D, PANSS), and patient population characteristics. Units include dosage (mg), frequency (times/day), and treatment duration (weeks, months).
Constraints on Knowledge Base Retrieval and Recall
Knowledge base retrieval and recall for psychiatric products face multiple challenges. First, data diversity and complexity require the knowledge base to integrate information from various sources and structures effectively. This is especially true for clinical research data, which needs fine-grained processing of numerical information like scale scores and statistical results.
Second, inconsistent update frequencies demand a flexible knowledge update mechanism. This ensures recalled drug information and clinical guidelines are current, preventing outdated or inaccurate advice. Documents contain many specialized terms and abbreviations, such as "SSRIs" and "SNRIs." The retrieval system needs strong semantic understanding to identify and link these terms. It must accurately recall information even when query terms do not exactly match the original text. Additionally, psychiatric diagnosis and treatment have high individual variability. The knowledge base must consider the contextual relevance of retrieval results to avoid over-generalization.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk size (Chunk Length) | 500–800 characters | Key information in psychiatric drug inserts and clinical guidelines often concentrates in longer paragraphs. Shorter chunks can break semantic connections. |
Chunk Overlap Length (Overlap Length) | 100–150 characters | Ensures complete capture of specialized terms or critical information at chunk boundaries, improving recall continuity. |
Recall count (Recall Count) | 8–12 items | Given the complexity of psychiatric diseases, increasing recall count can cover more potentially relevant information, aiding subsequent model judgment. |
Similarity threshold (Similarity Threshold) | 0.75–0.85 | Psychiatric terminology requires high precision. A relatively high threshold reduces irrelevant or ambiguous recalls. |
Rerank result count (Rerank Count) | 5 items | After initial recall, a reranking mechanism further refines the most relevant segments, reducing noise passed to the large language model. |
embedding_model | text-embedding-ada-002 or bge-large-zh-v1.5 | Selects models with strong semantic understanding and Chinese processing capabilities to handle the complexity of specialized psychiatric terminology. |
Common Pitfalls
- When querying a specific document in the workbench, the application may report that the document was not found. This can happen if the document was not correctly indexed in the knowledge base or if query keywords do not semantically match the document content.
- The knowledge base may hit its maximum capacity, preventing the addition of more knowledge. This usually results from underlying vector database (e.g., Milvus or PGVector) configuration limits or if the FastGPT configuration parameter
KNOWLEDGE_BASE_MAX_COUNTis not adjusted as needed. - Semantic or full-text search tools may return a "Connection error." This typically indicates a network connectivity issue, service not running, or misconfiguration of the vector database service (e.g.,
milvus) or full-text indexing service (e.g.,elasticsearch).
Verification Steps
- Perform a series of precise and fuzzy queries for core drug names and disease symptoms. Check if recall results include all expected key information and evaluate the completeness of the recalled content.
- Upload and index a document containing a new psychiatric drug insert. Verify the system can complete indexing within a reasonable time and accurately recall the document by drug name.
- Simulate user questions, such as "What are the side effects of clozapine?". Check if recall results accurately hit the adverse reactions section in the insert and evaluate the relevance threshold of the recalled segments.
- Regularly check the knowledge base management interface for index status and update timestamps. This ensures all critical documents are successfully processed and the update mechanism works as expected.
The values given 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.