What the Data for This Category Looks Like
Data for aerospace equipment-related financial reports primarily comes from public annual and semi-annual financial reports of listed companies in the aerospace equipment sector, as well as public industry statistical information released by national defense science, technology and industry authorities. Updates follow a fixed schedule: annual reports are released by April each year, semi-annual reports by August each year, and temporary announcements are released as needed. Document structures include standard financial statement modules, plus exclusive modules such as detailed R&D investment, military project contract execution status, equipment production capacity and equipment delivery progress descriptions. Fields include regular financial fields, as well as aerospace-specific fields: special contract amounts, equipment delivery quantity (units: units, sets, pieces), disclosed amounts for military confidential project compliance, and some fields use special measurement methods not found in general finance.
What Constraints These Characteristics Impose on Multi-turn Conversation and Prompt Engineering
The multi-source data of aerospace equipment financial reports requires distinguishing context boundaries between listed company public financial reports and industry statistical information during multi-turn conversations, to avoid mixing statistical standards from different data sources. Fixed update schedules require prompts to clearly limit use of only disclosed current-period report data, and prohibit introducing undisclosed internal information. Exclusive document modules require prompts in multi-turn conversations to guide the model to prioritize extracting special business fields such as special contracts and equipment delivery progress, rather than focusing solely on regular financial indicators. Special fields and units require verifying measurement consistency during multi-turn conversations; for example, equipment delivery quantities must use exclusive units such as units, sets, pieces, to avoid confusion with general financial amount units.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | 8000–12000 characters | Single aerospace equipment financial report documents often exceed 3000 characters. Multi-turn conversations need to retain context from multiple reports while avoiding exceeding model context limits |
rag_top_k | Top 6–8 results | Exclusive module content of aerospace equipment financial reports is scattered across different paragraphs. Sufficient relevant fragments must be retrieved to cover fields such as special contracts and equipment delivery progress |
prompt_template | "Please answer user questions based on the provided aerospace equipment financial report data, prioritize extracting exclusive fields such as special contract amounts and equipment delivery quantities, use corresponding units, and disclose confidential content in accordance with public disclosure standards if involved" | Adapts to exclusive business fields and disclosure rules of aerospace equipment financial reports, guides the model to focus on specific business content |
conversation_history_max_length | Top 3–5 turns | Multi-turn questions for aerospace equipment financial report analysis often focus on different modules of the same report. Excessive history will disrupt context focus on the current question |
file_parse_chunk_size | 1000–1500 characters | Exclusive module paragraphs of aerospace equipment financial reports are relatively long. Proper chunking retains the integrity of business context |
rag_similarity_threshold | 0.75–0.85 | General financial content unrelated to aerospace equipment financial reports must be filtered out, to accurately retrieve exclusive business fields |
The parameter values provided on this page are general recommendations used to establish configuration starting points. Actual values are affected by material format, data volume and business rules. Specific issues require case-by-case analysis. It is recommended to test on your own samples before finalizing settings.
Three Common Mistakes
- Incorrect offset parameter setting when calling the
get_conversation_listinterface, leading to abnormal numbers of returned data entries. Phenomenon: The conversation records returned by the interface do not match the expected range, for example, setting the offset to 10 but returning the first 10 records. Cause: The offset is not clearly defined as the starting index, it is not distinguished from the number of entries, and the logic of the offset and number of entries parameters is confused. - Global variable changes cannot take effect in the same conversation. Phenomenon: The first passed global variable value displays normally, but when attempting to modify the variable in subsequent conversations, the variable value is not updated. Cause: The dynamic update switch for session-level variables is not configured, variables are only loaded during session initialization, and real-time modification is not supported.
- Unable to retrieve user question content from each turn in multi-turn conversations. Phenomenon: When calling related components, only the current question can be obtained, and previous two user questions cannot be traced back. Cause: The context extraction configuration for conversation history is not enabled, and the component only reads the latest message of the current session.
How to Verify Proper Configuration
- Call the
get_conversation_listinterface, pass different offset parameters, and check whether the returned conversation record range matches expectations. - View the conversation history interface, confirm that each user question is correctly bound to the corresponding AI reply, with no misalignment.
- Initiate a multi-turn conversation, modify the preset global variable value, and check whether the variable value is updated synchronously in subsequent conversations.
- Call the conversation history extraction component, trace back three historical questions in sequence, and confirm that each user input content can be obtained normally.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-14.