Data Characteristics for This Category
Data for OA process initiation by an internal office assistant in the biopharmaceutical domain primarily comes from internal OA systems, ERP systems, and human resource management systems. This data updates frequently, especially on workdays, with real-time changes to process statuses and approval opinions. Document structures are typically a mix of structured and semi-structured data. For example, fields in a process application form are structured, while approval opinions or attachment content are semi-structured text. Fields include applicant, application time, process type, approval node, approval status, approval opinion, amount, department, and project number. The amount field may involve multiple currencies, and some fields may contain specific business codes or abbreviations.
Constraints Imposed by These Characteristics on Multi-Turn Conversations and Prompts
High-frequency data updates require the multi-turn conversation system to query the latest process status in real time, avoiding outdated information. The coexistence of structured and semi-structured data means prompt design must handle both field extraction and free-text understanding. This ensures the model accurately parses user intent, whether the user asks for a specific field value or a summary of approval opinions. The presence of multiple currencies and business codes means the model needs domain knowledge or knowledge base supplementation to correctly identify and process this information. For example, when a user queries "reimbursement process status," the system must differentiate between various reimbursement types and extract key information from complex business codes. Furthermore, process initiation involves multiple steps and approvers. Multi-turn conversations must guide the user through the entire process and provide relevant instructions at different nodes.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 800–1200 characters | Ensures coverage of key information like process initiation and approval opinions in multi-turn conversations, while controlling context length to reduce computational costs. |
similarityThreshold | 0.75 | Balances recall and precision. This ensures matching relevant process templates or approval rules and avoids interference from irrelevant information. |
top_k | Top 5 | Retrieves enough relevant process templates or approval rules during the knowledge base retrieval phase to improve the accuracy of subsequent model understanding. |
reRankTopN | 3 | Re-ranks retrieved results, ensuring the most relevant process information or approval rules are prioritized. |
promptTemplate | Includes {{context}} and {{query}} | Guides the model to understand user intent and generate business-logic-compliant responses based on process data. |
history_len | 5 turns | Maintains sufficient conversation history. This allows the model to understand contextual dependencies in multi-turn conversations, especially the continuity of process approvals. |
Three Common Mistakes
- Inaccurate process status or approval results returned by the model. This can occur if process definitions or approval rules in the knowledge base are not updated promptly, leading the model to cite old data.
- The model fails to correctly identify and provide information when users query a specific process number. This happens when the prompt does not effectively guide the model to extract key entities like process numbers from complex semi-structured text.
- In multi-turn conversations, after a user initiates a process, the model fails to guide the user to fill in necessary fields. This manifests as conversation interruption or the model repeatedly asking for already provided information. This usually happens when
history_lenis set too low, causing the model to forget previous context.
How to Confirm Proper Configuration
- Simulate various OA process initiation scenarios, including normal, abnormal, and rejected processes. Check if the model accurately understands user intent and provides correct process guidance.
- Test process queries of varying complexity. For example, query all pending processes for a specific applicant, or query the total reimbursement amount for a department within a specific period. Check if the model can accurately extract and summarize information from structured data.
- Observe the model's response to changes in process status during multi-turn conversations. For example, when a process changes from "pending approval" to "approved," can the model update its response in real time? Set an acceptable delay threshold based on business requirements.
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.