Knowledge Base Retrieval and Recall for Rational Drug Use

Data Characteristics

Data for rational drug use regulations primarily comes from policy documents, technical guidelines, and drug catalogs issued by national health and medical insurance commissions. It also includes treatment norms, formularies, and pharmaceutical management systems developed internally by hospitals. These documents are typically in PDF, Word, or structured database formats. Data update frequency is relatively stable; policy and regulation updates occur several times a year, while drug catalogs may be adjusted quarterly or semi-annually. Documents are often chapter-based, containing numerous technical terms, drug names, dosage units, indications, and contraindications. Key information, such as generic names, brand names, formulations, specifications, routes of administration, and dosages, is frequently presented in tables or lists.

Constraints on Knowledge Base Retrieval and Recall

Rational drug use documents are highly structured, but their language is precise and specialized, demanding high accuracy in semantic understanding and exact matching. Synonyms, abbreviations, differences between drug formulations, and accurate recognition of dosage units (e.g., mg, g, IU) directly impact the usability of retrieval results. Although update frequency is not high, each update can involve critical revisions or additions, requiring the knowledge base to respond and update its index quickly. Furthermore, due to patient safety concerns, the accuracy and completeness of retrieval results are crucial. The system must avoid misjudgments caused by incomplete recall or interference from irrelevant information. A core challenge is associating and recalling multiple entities in a query (e.g., "a certain drug" + "a certain disease" + "dosage").

Configuration Settings

Configuration Item	Recommended Value	Rationale
`Chunk size` (Chunk Size)	`300–500 characters`	Regulatory texts are highly logical. Chunks that are too short may split complete clauses, while chunks that are too long increase noise.
`Chunk overlap` (Chunk Overlap)	`50 characters`	Ensures contextual continuity between paragraphs, preventing critical information from being cut off.
`Recall count` (Recall Count)	`8–12 items`	Guarantees multi-faceted coverage of query intent while balancing the model's processing capacity.
`Similarity threshold` (Similarity Threshold)	`Calibrated by actual measurement`	Requires balancing accuracy and recall rate based on actual query performance.
`Rerank result count` (Reranked Return Count)	`5 items`	Focuses on the most relevant results, reducing the model's inference burden.
`maxContext`	`4096 tokens`	Ensures the model can process a sufficiently long context to fully understand regulatory clauses.

Common Pitfalls

The knowledge base fails to answer questions related to rational drug use regulations, resulting in empty responses. This often occurs due to an improper document chunking strategy, leading to critical information being split, or a Similarity threshold (Similarity Threshold) set too high, failing to recall relevant text.
The answer cites irrelevant policy provisions or drug instructions. This happens when Recall count (Recall Count) is too high, or vector retrieval's semantic understanding is insufficient, failing to effectively distinguish similar but non-target content.
The API call returns a response missing knowledge base citations. This may mean the use_knowledge_base parameter in the API request is not set to True, or there is an issue with the data flow configuration between the knowledge base and the model, preventing knowledge base retrieval results from being passed to the model for generation.

Verification

Query typical rational drug use scenarios (e.g., "dosage adjustment for a certain drug in patients with renal insufficiency"). Check if the recall results include all relevant policy provisions and drug insert content, and confirm their completeness.
Simulate updated regulations by re-uploading and indexing them. Then, query relevant content to verify if the knowledge base can reflect the latest changes promptly.
Use FastGPT's debugging interface to observe the Recall count (Recall Count), Similarity Score, and the final model-generated answer for each query. Evaluate the accuracy of the answer and the reliability of the cited sources.

Note: The values provided are common starting points. It is crucial to measure their effectiveness 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.