Knowledge Base Retrieval and Recall for Patient Assistance

Data Characteristics

Patient Assistance Program (PAP) data originates from pharmaceutical companies, charities, and medical institutions. Data update frequency typically aligns with program policy adjustments, drug batch changes, and patient application, approval, and dispensing processes. Updates may occur monthly or quarterly. Document formats vary, including application forms, patient informed consent forms, drug inserts, program management guidelines, Frequently Asked Questions (FAQs), and various policy notices. Drug inserts contain standard medical fields such as active ingredients, indications, dosage, and adverse reactions. Program guidelines involve specific business fields like application conditions, review processes, drug reimbursement ratios, and assistance periods, often accompanied by timestamps and version control information like effective dates, expiry dates, and version numbers.

Constraints Imposed by These Characteristics on Knowledge Base Retrieval and Recall

The multi-source nature and update frequency of PAP data require the knowledge base to support efficient file ingestion and version management to ensure the real-time nature and accuracy of retrieved content. Complex document structures, encompassing both standardized drug information and unstructured policy interpretations, demand that the knowledge base effectively identifies different information types during segmentation and maintains their semantic integrity. For example, drug adverse reactions and program application conditions require distinct handling during retrieval. Furthermore, the medical and business terminology in the data places higher demands on tokenization and vectorization models, which must accurately capture the meaning of specialized vocabulary. Numerical fields in policy terms, such as reimbursement ratios or assistance periods, require precise matching during retrieval to avoid incorrect recall due to numerical approximation.

Configuration Recommendations

Configuration Item	Recommended Value	Rationale
`Chunk size`	`800–1200 characters`	Balances contextual completeness and retrieval efficiency, ensuring each segment contains sufficient policy or drug information.
`Chunk Overlap Length`	`100–200 characters`	Ensures continuity of information across segments, preventing important information from being truncated.
`Recall count`	`Top 5–8 entries`	Covers multiple highly relevant document snippets, increasing information coverage.
`Similarity threshold`	`0.7–0.85`	Balances retrieval precision and recall rate, avoiding interference from irrelevant content.
`Rerank result count`	`Top 3 entries`	Further filters the most relevant core information based on initial recall.
`PARSE_FILE_TIMEOUT_SECONDS`	`600 seconds`	Accommodates the time required to parse large policy documents or drug inserts.

Common Pitfalls

Retrieval results include outdated policies or invalid program information. This occurs when the knowledge base is not updated promptly or document effective/expiry dates are not correctly marked.
When a user asks about drug usage, the system recalls program application procedures. This indicates a significant deviation between recalled content and user intent. This may be due to the tokenizer failing to effectively distinguish semantic boundaries between drug knowledge and program management knowledge, or the vectorization model having insufficient differentiation capability for different text types.
For questions about specific drug dosages or reimbursement ratios, the answer contains inaccurate or missing numerical values. This manifests as answers that do not align with facts. This often occurs when the knowledge base fails to correctly parse numbers or units during structured data ingestion, leading to a lack of precise matching during retrieval.

Validation of Configuration

Select various types of patient assistance program questions, such as drug side effects, application conditions, and reimbursement ratios. Check if the chunk_id in the retrieval results covers the key parts of the expected documents.
For time-sensitive questions, such as "latest policy for a certain program," verify the version or effective_date fields of the recalled documents to ensure the latest version is returned.
Randomly select a batch of common user questions. Conduct simulated dialogue tests to observe whether the knowledge base segments cited in the AI's answer directly support the content of the answer, and check if the cited source field points to the correct file.

The values provided 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.