Data Characteristics of This Category
Market access R&D documents typically include drug registration approvals, medical device registration certificates, clinical trial reports, pharmaceutical research reports, non-clinical research reports, package inserts, and labels. These documents are often in PDF or scanned image format, featuring complex structures with numerous tables, graphs, and specialized terminology. Update frequency depends on policy regulations, product iterations, and clinical data updates, usually quarterly or annually, though some critical policies may be released at any time. Fields involved include generic drug name, active ingredients, indications, dosage and administration, adverse reactions, manufacturer, approval number, and expiration date. Units cover dosage (mg, g), concentration (%), time (h, min), and volume (mL), often accompanied by abbreviations and symbols.
Constraints Imposed by These Characteristics on "Multiturn Conversations and Prompts"
The complex and diverse document structures demand high accuracy in information extraction. Multiturn conversations require precise identification of user intent, such as querying a specific side effect of a drug or comparing approval numbers of different devices. This necessitates prompts that guide the model to locate key information within unstructured text. The presence of specialized terminology and abbreviations means prompt design must consider context to avoid misunderstanding or omission. The irregular nature of document updates, especially policy changes, requires the knowledge base to synchronize quickly and reflect the latest information in multiturn conversations, preventing outdated advice. Additionally, users may upload image-based documents for queries, requiring the system to have image recognition capabilities to convert them into processable text for structural analysis and dialogue.
Configuration Settings
| Configuration Item | Suggested Value | Rationale for This Value |
|---|---|---|
maxContext | 8000 tokens | Accommodates the context length of complex documents, ensuring coherence in multiturn conversations. |
Chunk size (Chunk Length) | 500 characters (characters) | Balances recall granularity and chunk integrity, minimizing information truncation. |
Recall count (Recall Count) | Top 8 entries (top 8) | Covers a broader range of potentially relevant information for complex queries. |
Similarity threshold (Similarity Threshold) | 0.75 | Precisely matches specialized terminology, avoiding generalized recall. |
Rerank result count (Reranked Return Count) | 3 entries (3 items) | Highlights the most relevant results, improving response efficiency in multiturn conversations. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds (seconds) | Handles parsing large PDF documents, preventing timeout interruptions. |
Three Common Mistakes
- In a multiturn conversation, a user asks about a drug with a specific approval number, but the system returns information for a drug with the same name but a different approval number. This happens because the prompt insufficiently weights the "approval number" as a critical identification field, causing the model to prioritize matching the drug name.
- A user uploads a scanned market access document, but the dialogue system fails to extract any information, returning an empty response. This occurs because the system is not configured for or has not correctly enabled Optical Character Recognition (OCR), making it unable to process image input.
- When querying a specific parameter of a device, the dialogue system repeatedly asks the user "which device," even if the user clearly mentioned it in the previous turn. This indicates that the multiturn conversation's context management mechanism does not effectively retain entity information from prior turns.
How to Confirm Correct Configuration
- Select 5 representative market access documents, including tables, graphs, and specialized terminology. Conduct multiturn questioning tests to check the extraction accuracy of key information (e.g., approval number, indications, adverse reactions).
- Upload a recently updated market access policy document. Test whether the system can cite the latest policy terms in multiturn conversations and differentiate them from older policies, verifying the timeliness of knowledge updates.
- Prepare 3 sets of queries containing specialized abbreviations and synonyms. Test whether the system correctly understands user intent and provides relevant answers, evaluating the prompt's depth of understanding for specialized terminology.
- Upload a scanned document containing a complex table. Ask for the value of a specific field in the table to confirm that the system can accurately extract data through OCR and structural analysis.
Note that the values provided are common starting points and should be measured against specific 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.