Vector Models and Indexing for Smart Triage Products

Data Characteristics in This Category

Smart triage data primarily comes from authoritative medical institutions. This includes disease diagnosis and treatment guidelines, drug inserts, medical literature, clinical pathways, and structured/unstructured text from physician consultation experiences. Data updates are generally stable. However, events like new drug approvals, updated treatment protocols, and epidemiological changes trigger rapid localized updates. Document structures are diverse, including plain text guidelines, drug inserts with tables and diagrams, and FAQ-style question-and-answer sets. Fields and units are highly specialized. Examples include disease names, symptom descriptions, examination results (e.g., white blood cell count in a complete blood count report with units of 10^9/L), drug ingredients, dosages (e.g., mg, g), usage instructions, and contraindications.

Constraints Imposed by These Characteristics on Vector Models and Indexing

The highly specialized and diverse nature of smart triage data requires vector models to accurately capture the semantic relationships of medical terminology. Models must also distinguish subtle differences between similar symptom descriptions. For instance, abdominal pain and stomach pain might seem similar in everyday language but refer to different organs medically. Diverse document structures, especially drug inserts with tables and diagrams, challenge data preprocessing and chunking strategies. This requires ensuring important information remains intact. Stable update frequency with localized rapid update needs means the index must support efficient incremental updates, avoiding full rebuilds. The specialized nature of fields and units requires standardization and entity recognition before vectorization. This reduces ambiguity and ensures retrieval precision. For example, when a user asks about the dosage of ibuprofen, the system should prioritize retrieving document segments containing dosage information.

Configuration Guidelines

Configuration Item	Recommended Value	Rationale
`Chunk Length`	`500–800 characters`	Balances semantic completeness with recall efficiency. Avoids overly long chunks that cause information redundancy and overly short chunks that lead to semantic discontinuity.
`Chunk Overlap Length`	`100 characters`	Ensures continuity of context at chunk boundaries, improving retrieval recall.
`Recall Count`	`Top 10–20`	Provides sufficient candidates for the subsequent reranking stage while ensuring retrieval coverage.
`Similarity Threshold`	Calibrate based on actual measurements	Requires testing with specific vector models and corpora to balance recall and precision.
`Rerank Return Count`	`Top 3–5`	Focuses on the most relevant results users likely need, reducing user filtering burden.
`Max Index File Size`	`2048 MB`	Prevents individual index files from becoming too large, which can affect update and query performance.

Three Common Pitfalls

Knowledge base query results are empty or irrelevant: This often happens when the vector model cannot understand medical terminology or abbreviations in the user's query, leading to inaccurate vector representations.
Training orders remain pending for a long time: This usually results from insufficient underlying vector database resources or a backlog in the training queue, affecting the timeliness of knowledge updates.
Retrieved drug insert images cannot be displayed: This occurs because image content was not effectively vectorized or indexed. Retrieval then only recalls text information, lacking visual aids.

How to Verify Configuration

Simulate user queries. Check if recalled results contain key entity information like diseases, symptoms, and drugs. Verify the relevance score threshold.
Regularly use the GET /api/v1/kb/collections/{collectionId}/stats API to check the knowledge base index status. Ensure indexedCount matches the actual data volume.
Perform retrieval tests with complex documents, including tables and images. Verify correct recall and display of all relevant information. For example, validate examination items and results fields in diagnostic reports.

Note: The values provided are common starting points. Measure performance against your own samples to determine optimal configurations.

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.