Multiturn Conversation and Prompts for Internal Meeting

Data Characteristics

Meeting minutes in the biopharmaceutical domain are typically unstructured text. Sources include internal meeting records, project workshop minutes, and clinical trial protocol review meeting minutes. These documents are usually stored in internal document management systems or collaboration platforms. Update frequency depends on meeting frequency, ranging from daily to weekly or monthly. A meeting minute document typically contains core fields such as meeting topic, time, location, attendees, discussion points, key decisions, action items, and responsible parties. Discussion points and key decisions are often free-text, involving specialized terminology, experimental data references (e.g., dose units mg/kg, time units h or day, concentration units μM), and project numbers.

Constraints on Multiturn Conversation and Prompts

The diversity and unstructured nature of meeting minute data impose specific requirements on multiturn conversation and prompt design. Due to the extensive use of specialized terminology and abbreviations, prompts must guide the model to accurately understand context, preventing semantic deviations caused by inaccurate recognition of professional vocabulary. The update frequency of meeting minutes requires the knowledge base to quickly index and retrieve the latest content. In multiturn conversations, user queries may involve decisions from the most recent meeting, so the system must recall the most current data. Meeting minutes also contain action items and responsible party fields, which means the model needs to extract structured information in multiturn conversations to answer queries like "who is responsible for completing a task." For numerical values and units, prompts must reinforce unit recognition and matching to avoid errors where values and units are disconnected.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`Chunk size`	`800–1200 characters`	Retains the context of a complete topic or decision within meeting minutes, while preventing excessively long segments from impacting recall efficiency.
`Recall count`	`Top 5–8 entries`	Covers key information points potentially involved in user queries, balancing recall precision with model processing load.
`Similarity threshold`	`0.75–0.85`	Ensures recalled meeting minute segments are highly relevant to user queries, filtering out low-quality matches.
`Rerank result count`	`Top 3 entries`	Further refines the most relevant segments from recall results, improving model answer accuracy.
`maxContext`	`32000 token`	Supports longer multiturn conversation history and complex meeting minute content, preventing context loss.
`PARSE_FILE_TIMEOUT_SECONDS`	`300 seconds`	Allows sufficient time for the system to process large meeting minute documents, preventing parsing timeouts.

Common Pitfalls

The model fails to accurately identify project numbers or drug names in meeting minutes during a conversation, leading to responses that do not match the actual query. This occurs because the knowledge base lacks sufficient weighting or preprocessing for these specific entities during vectorization.
A user asks about a decision from the most recent meeting, but the model cites content from minutes several months old. This occurs because the knowledge base's indexing update mechanism fails to synchronize newly uploaded meeting minute files in a timely manner.
Tool call results do not display in the chat window, or the returned tool call result is empty. This may occur if the schema definition in the tool call configuration does not match the actual tool's output format, leading to parsing failure.

Verification Steps

Upload a meeting minute document containing the latest decisions. Then, conduct multiturn conversation tests to verify if the model can accurately cite key decisions from that document and indicate the source of information.
Construct complex queries containing specialized terminology and dosage units. Check if the model can correctly understand and provide answers with the correct units.
Use the "Debug" feature in the FastGPT backend to view the input and output of each tool call in multiturn conversations. Confirm that tool input parameters are correctly passed and that the json structure of the output results meets expectations.
For meeting minutes containing action items and responsible parties, ask "Who is responsible for completing [a certain task]?" Verify if the model can accurately extract and return the corresponding responsible person's name.

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.