Knowledge Base Retrieval and Recall for DTP Pharmacy

Data Characteristics

DTP pharmacies use data for clinical trial pre-screening primarily from pharmaceutical companies. This includes clinical trial protocols, patient recruitment criteria, drug inserts, investigator brochures (IBs), and internal patient medical summaries. This data updates frequently, typically monthly or quarterly, due to new drug approvals, clinical trial progress, or protocol amendments. Documents are usually unstructured PDF text, containing extensive medical terminology, dosage units (e.g., mg/kg, IU), and specific fields (e.g., Inclusion Criteria, Exclusion Criteria, Adverse Events). Some data may appear as structured tables within PDF reports, describing patient characteristics or treatment outcomes.

Constraints Imposed by Data Characteristics on Knowledge Base Retrieval and Recall

The data characteristics of DTP pharmacy clinical trial pre-screening impose multiple constraints on knowledge base retrieval and recall. First, the semantic complexity of unstructured medical text requires RAG systems to have strong semantic understanding, accurately identifying medical entities and relationships. Second, frequent data updates necessitate an efficient incremental update mechanism for the knowledge base, preventing the recall of outdated or incorrect trial information. Third, structured information within nested tables in PDF documents is easily lost during traditional text chunking, affecting precise matching of patient characteristics. Furthermore, strict inclusion and exclusion criteria demand that the system accurately matches multiple conditions, such as patient age, disease stage, and concomitant medications. Any subtle deviation can lead to inaccurate pre-screening results. Identifying dosage units and specific fields, such as mg/kg or Adverse Events, is crucial for ensuring retrieval accuracy.

Configuration Guidelines

Configuration Item	Recommended Value	Rationale
`Chunk Size`	`500-800 characters`	Balances context completeness for medical text and retrieval efficiency, avoiding irrelevant information from overly long chunks.
`Chunk Overlap`	`50-100 characters`	Ensures critical information at chunk boundaries is not lost, improving recall robustness.
`Recall Count`	`8-12 items`	Balances coverage with computational load for subsequent re-ranking and generation stages, aiming for response times under `30 seconds`.
`Similarity Threshold`	Calibrate by measurement	Determine through small-sample testing with specific datasets and business requirements, typically between `0.75-0.85`.
`Rerank Count`	`3-5 items`	Further refines recall results, improving final answer accuracy and reducing unnecessary computational overhead.
`PARSE_FILE_TIMEOUT_SECONDS`	`600 seconds`	Addresses potentially long parsing times for large clinical trial protocol PDF files, preventing timeout.

Common Pitfalls

Symptom: API calls to the knowledge base return content without knowledge base information. Reason: The API call did not correctly pass chat_id or app_id, preventing the request from being associated with the specified knowledge base.
Symptom: After adding knowledge base content, retrieval response time significantly slows, exceeding 30 seconds. Reason: Recall Count or Rerank Count is set too high, leading to excessive computation during retrieval and re-ranking.
Symptom: DTP pharmacy clinical trial pre-screening results incorrectly include patients who do not meet the criteria. Reason: The chunking strategy failed to effectively preserve tabular inclusion/exclusion criteria within PDF documents, resulting in the loss of critical structured information during retrieval.

Validation Steps

Upload typical clinical trial protocol PDF files through the FastGPT management interface. Check the chunk preview after parsing to ensure critical information (e.g., Inclusion Criteria, Exclusion Criteria) is segmented completely and independently.
Simulate real patient inquiries. Input multi-turn conversations related to clinical trial pre-screening. Observe whether the system's recalled content accurately covers patient characteristics and trial criteria. Compare with the original document to confirm the Similarity metric of the recall.
Use FastGPT's API interface for batch testing. Record response times under different Recall Count and Rerank Count configurations. Determine thresholds based on acceptable business latency.
For queries containing dosage units (e.g., mg/kg) and specific fields (e.g., Adverse Events), verify that recall results accurately identify and match this information. Adjust tokenizer configurations or vector models if necessary.

Note: The values provided are common starting points. Measure against specific samples to determine optimal configurations for individual use cases.

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.