Multi-turn Conversation and Prompts for IVD Diagnostic

Data Characteristics for this Category

IVD diagnostic reagent pharmacovigilance data originates primarily from clinical trial reports, real-world studies, post-market adverse event surveillance reports, and regulatory risk communications. This data exists in both structured formats (e.g., CIOMS I forms, MedWatch forms) and unstructured formats (e.g., handwritten clinician notes, patient feedback text). Data updates frequently, especially during post-market surveillance, where new adverse event reports may arise daily. Document structures are complex, potentially including product inserts, batch information, patient demographics, diagnostic results, adverse event descriptions, and management actions. Specifically, IVD diagnostic reagent data involves unique fields such as batch number, expiration date, and detection principle. Units may include biochemical units like IU/mL, ng/dL, and optical density units like OD value.

Constraints Imposed by These Characteristics on "Multi-turn Conversation and Prompts"

The high update frequency of IVD diagnostic reagent data requires FastGPT's knowledge base to quickly synchronize the latest information. This ensures the timeliness of adverse event information referenced in multi-turn conversations. The coexistence of structured and unstructured data means prompt design must account for both precise extraction from structured fields and semantic understanding of unstructured text. For example, if a user queries adverse reactions for a specific reagent batch, the system must identify the batch number from unstructured descriptions and combine it with structured data for matching. Unique biochemical and optical density units demand that the model accurately recognize and convert these specialized terms. This prevents misjudgments due to unit confusion during multi-turn follow-up questions. Furthermore, multi-turn conversations require context memory to handle referential questions like "other adverse reactions for that batch," allowing the system to locate key information within complex reports.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`maxContext`	`6`	Balances context retention and computational cost, ensuring critical information is not lost.
`Chunk size` (Segment Length)	`800–1200 characters`	Accommodates longer descriptive texts in adverse event reports, improving recall quality.
`Recall count` (Recall Count)	`Top 5 entries`	Covers multiple potentially relevant adverse event reports, increasing hit rate.
`Similarity threshold` (Similarity Threshold)	`0.78`	Excludes noisy information, focusing on diagnostic reagent events highly relevant to the query.
`Rerank result count` (Rerank Return Count)	`Top 3 entries`	Further refines results, prioritizing the most relevant core information.
`ENABLE_FILE_UPLOAD`	`true`	Supports uploading adverse event report PDFs or images for analysis.

Three Common Mistakes

During multi-turn conversations, a user mentions "the reagent kit discussed last time," but the system fails to link it to specific product information. This occurs because maxContext is set too low, leading to insufficient context memory.
Uploaded PDF-format adverse event reports are not parsed correctly, resulting in empty content. This might be due to the PARSE_FILE_TIMEOUT_SECONDS parameter being set too short, which is insufficient for processing complex documents.
When querying adverse reactions related to specific units (e.g., IU/mL), the system returns results with mismatched or confused units. This happens because the prompt does not explicitly instruct the model to standardize professional units.

How to Confirm Proper Configuration

When testing multi-turn conversations, mention a batch number or product name discussed in previous turns during an intermediate turn. Observe if the system correctly identifies it and continues the conversation.
Upload an IVD adverse event report PDF containing complex charts and specialized terminology. Verify that its content, especially the adverse event description field, is extracted completely and accurately.
Construct queries containing different units (e.g., ng/dL and µg/L). Check if the model's recognition and conversion of units are consistent across multi-turn conversations. Adjust unit handling instructions in the prompts based on actual needs.

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.