Citation and Traceability for Home Medical Device

Data Characteristics for This Category

Pharmacovigilance data for home medical devices primarily comes from user feedback, after-sales records, recall notifications, regulatory reports, and post-market studies. This data often updates frequently, especially user feedback and after-sales records, which can be generated in real-time. Document formats vary, including unstructured user descriptions, structured adverse event report forms, product manuals, usage guidelines, and technical specifications. Common fields include product model, batch number, manufacturing date, adverse event description, occurrence time, severity, treatment measures, patient age, gender, and medication history. Adverse event descriptions can contain extensive free text, while product information is typically standardized coding.

Constraints Imposed by These Characteristics on Citation and Traceability

The highly unstructured nature of home medical device data, particularly free text in user feedback, presents challenges for citation and traceability. The length and complexity of original text require knowledge bases to maintain semantic integrity during chunking, preventing critical information from being fragmented. The real-time nature of data updates means knowledge base content needs frequent synchronization or real-time retrieval capabilities to ensure citation timeliness. Additionally, due to the importance of fields like product model and batch number, indexing must ensure these key identifiers are accurately recognized and associated to support precise traceability. The diversity of document formats, such as PDF, Word, and images, demands robust multi-format parsing capabilities from the knowledge base, and the ability to render citations in a format close to the original document to enhance user experience and information credibility.

Configuration Settings

Configuration Item	Suggested Value	Rationale
`Chunk size` (Chunk Length)	`800–1200 characters`	Balances semantic integrity of free-text user feedback, avoiding context loss from overly short chunks.
`Recall count` (Recall Count)	`Top 10 entries`	Given the complexity of adverse event descriptions, increases recall count to cover potentially relevant information.
`Similarity threshold` (Similarity Threshold)	`0.75–0.85`	Ensures recalled results are highly relevant to user queries, reducing interference from irrelevant information.
`Rerank result count` (Reranked Return Count)	`Top 5 entries`	Prioritizes the most relevant citation sources, improving efficiency in obtaining key information.
`PARSE_FILE_TIMEOUT_SECONDS`	`600 seconds`	Accommodates parsing time for large product manuals or complex adverse event reports, preventing timeouts.
`CONTEXT_RENDER_FORMAT`	`Markdown`	Enhances readability of cited content, allowing engineers to quickly understand document structure.

Common Misconfigurations

Cited source content does not render as Markdown as expected, appearing as raw text. This can occur if the Markdown renderer is not correctly enabled or recognized in the knowledge base configuration.
The sourceid field of a citation is empty or points to an incorrect database record. This manifests as traceability information not linking to the original document. This typically happens when critical metadata fields are not correctly parsed or stored during file upload.
When a user asks consecutive questions, subsequent questions show the exact same citation sources as the previous question, even if the question topic has shifted. This can be due to an excessively large maxContext setting, causing the system to over-rely on historical context and fail to update its recall strategy in a timely manner.

How to Verify Correct Configuration

Submit a PDF-formatted adverse event report containing complex tables or lists. Check if citation sources correctly parse and render the table or list structure.
Upload an after-sales record containing a product model and batch number. Then, ask questions related to that product model and batch number. Verify if the sourceid in the citation accurately points to the original record.
Test the same type of document uploaded at different times. Observe the timeliness of citation sources, ensuring the system prioritizes recalling the latest version or most relevant documents, and check if their timestamps are accurate.
Simulate a multi-turn conversation with a user. Observe changes in citation sources across different turns. Determine if they update appropriately as the conversation topic evolves, and check if Recall count (Recall Count) and Rerank result count (Reranked Return Count) meet expectations.

The values provided are common starting points. Measure them against your own samples to determine the most suitable configuration.

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.