Multiturn Conversation and Prompts for Automated Management

Data Characteristics for This Category

Archiving data primarily originates from WeChat Work group chat messages (text, images, files), meeting minutes, experimental reports, and clinical data. These can be structured or semi-structured documents. Data update frequency varies based on group activity and business needs, ranging from multiple times daily to several times weekly. Document structures are diverse. Group chat messages are typically short texts, including timestamps, sender, and message content. Meeting minutes may contain topics, discussions, conclusions, and action items. Experimental reports often have fixed fields such as experiment objective, methods, results, and analysis. Common fields and units include date (yyyy-MM-dd HH:mm:ss), personnel names, project numbers, drug batch numbers, dosages (mg, ml), and experimental data (mol/L, ng/mL). Accurate identification and extraction of this information are essential for archiving.

Constraints Imposed by These Characteristics on Multiturn Conversation and Prompts

The short text nature and high update frequency of WeChat Work group messages require a multiturn conversation system to respond quickly and process a large volume of fragmented information. The continuous accumulation of conversation history challenges context management, requiring effective identification of historical conversations relevant to the current query, avoiding interference from irrelevant information. The semi-structured nature of documents like meeting minutes and experimental reports means prompt design must balance information extraction and content summarization, not relying solely on keyword matching. The specialized terminology and diverse units in the biomedical field demand high model understanding and generation accuracy. Prompts must guide the model to focus on identifying specific fields and units, for example, distinguishing mg from μg. Additionally, multiturn conversations may involve tracing and comparing historical archived records, setting higher standards for RAG retrieval accuracy and recall efficiency.

Configuration Guidelines

Configuration Item	Recommended Value	Rationale
`maxContext`	`2000 characters`	Balances WeChat Work short messages and some meeting minute lengths, maintaining context relevance.
`Chunk size` (Chunk Length)	`500 characters`	Optimizes RAG recall accuracy, preventing single chunks from being too long and diluting core information.
`Recall count` (Recall Count)	`Top 8`	Covers multiple relevant records that may be involved in multiturn conversations, improving relevance.
`Similarity threshold` (Similarity Threshold)	`0.75`	Ensures highly relevant retrieval results, reducing interference from low-quality or irrelevant information.
`Rerank result count` (Rerank Return Count)	`Top 5`	Refines the final context presented to the model while maintaining relevance.
`PARSE_FILE_TIMEOUT_SECONDS`	`600 seconds`	Addresses the parsing needs for potentially large or complex structured documents in the biomedical field.

Three Common Mistakes

After uploading documents, RAG retrieval results may lack images or specific structured data. This occurs when the document parser does not correctly handle image links or specific fields, leading to their exclusion during RAG indexing.
During multiturn conversations, the model frequently mentions irrelevant historical information in its responses. This happens if maxContext is set too high or Similarity threshold (Similarity Threshold) is too low, causing the model to include excessive irrelevant context when generating replies.
Specific drug dosage units (e.g., mg and g) mentioned in conversations are confused or incorrectly converted. This occurs if prompts do not explicitly instruct the model to focus on unit identification, or if the model's training data lacks sufficient examples for multi-unit recognition.

How to Verify Configuration

After uploading a Markdown document containing image links, verify through RAG retrieval that image links are correctly indexed and returned.
Simulate multiturn conversations, querying specific information from historical records. Cross-reference the model's responses with key information (e.g., date, project number) from the original archived records to confirm consistency.
For queries involving specialized terminology and measurement units, check if the corresponding terms and units in the model's response are accurate. The deviation from expected results should be within an acceptable range.
Observe conversation logs to confirm that the Recall count (Recall Count) and Rerank result count (Rerank Return Count) for each RAG retrieval match the configured settings, and that the recalled content is highly relevant to the current query.

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.