Multi-Turn Conversations and Prompts for CDMO Products

CDMO Data Characteristics

Core data for Contract Development and Manufacturing Organization (CDMO) products comes from project contracts, technology transfer agreements, batch production records, analytical method validation reports, and quality control documents. Data updates align closely with project cycles, typically occurring at project milestones or after batch production. Document structures are a mix of structured and semi-structured formats. For example, batch production records are often tabular, containing fields like material batch number, input quantity, operational steps, and Critical Process Parameters (CPP). Analytical reports may include instrument models, detection methods, Result Value, and Unit. Project change records and deviation handling reports also provide important supplementary data, with less frequent updates.

Constraints on Multi-Turn Conversations and Prompts

The highly structured and time-sensitive nature of CDMO data requires precise matching of specific project, batch, or phase information in multi-turn conversations. For instance, when a user queries key process parameters for a specific batch, the system must accurately locate that batch's production record and extract the CPP field value. The semi-structured nature of data means prompt design must handle both free-text queries and structured field queries. Frequent data updates demand efficient index update mechanisms for the knowledge base, ensuring real-time accuracy of conversation results. Furthermore, data isolation and access control between different projects are critical to prevent information leakage, influencing user authentication and data filtering logic before retrieval in the conversation system.

Configuration Settings

Configuration Item	Suggested Value	Rationale
`maxContext`	`2000 characters`	Balances context length and model inference cost, covering general multi-turn conversation scenarios.
`Recall Count`	`Top 5`	Prioritizes precision of recall results, reducing interference from irrelevant information.
`Similarity Threshold`	`0.75`	Addresses the precise matching requirements for CDMO terminology, improving recall relevance.
`Rerank Return Count`	`3 items`	Further filters recalled results, enhancing the quality of the final presented content.
`System Prompt`	See explanation below	Guides model behavior, ensuring output format and content meet professional requirements.
`Segment Length`	`500 characters`	Adapts to the common length of technical description paragraphs in CDMO documents.

Common Mistakes

After a user question, results appear as links in a sidebar instead of directly in the chatbox. This usually happens when the AI conversation configuration does not embed knowledge base retrieval results directly into the reply, but instead provides reference links.
The system cannot maintain context after each conversation, requiring re-asking. This may be due to a maxContext parameter set too low, preventing the model from retaining sufficient historical conversation information.
AI responses consistently include Markdown formatting, even when not desired. This typically occurs when the System Prompt contains instructions that force Markdown output; the prompt needs modification to remove these requirements.

How to Verify Configuration

For specific projects or batches, simulate multi-turn questions to check if the conversation accurately maintains context and correctly references historical information.
Test queries for different types of CDMO data (structured, semi-structured) to observe if AI responses accurately extract key fields like Result Value and Unit.
Check the format of AI output results to ensure they meet expectations, for example, by not including unnecessary Markdown tags, or by presenting tabular data as required.

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.