What Data for This Category Looks Like
Data for this category primarily comes from periodic disclosure announcements of domestic and overseas listed entities, and publicly available industry operation briefings from industry associations. Annual reports are published before April 30 of the following year. Interim reports are published before August 31. Quarterly reports are published within one month after the end of the corresponding quarter. Most documents are in PDF format, with some accompanying structured supplementary documents. These include modules such as core operating data, product category revenue, raw material procurement costs, R&D investment amounts, cash flow and liability details. Fields include product sales volume (unit: ton), revenue amount (unit: ten thousand yuan / hundred million yuan), unit production cost (unit: yuan / kilogram), procurement unit price (unit: yuan / ton), and more.
What Constraints These Characteristics Impose on Multi-turn Dialogue and Prompting
Different listed entities use different disclosure standards. Some enterprises do not separately split revenue data for paint and ink businesses. Multi-turn dialogue must first guide users to clarify target enterprises and disclosure standards, to avoid data confusion. Financial report update cycles are fixed and span long periods. Multi-turn dialogue must support users switching analysis across different report periods, and retain report period context from historical conversations. Financial report documents include multiple modules of segmented data. Multi-turn dialogue must support gradual breakdown of requirements, such as first confirming analysis of revenue composition before deeply inquiring about raw material cost changes. Field units vary across enterprises. Prompts must include unit calibration rules to automatically unify data units in conversations, preventing calculation or comparison errors.
How to Set Configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxHistory | First 6 chat records | Multi-turn dialogue for paint and ink financial reports typically requires retaining three core contexts: report period, enterprise standards, and core data. 6 records cover complete conversation logic without occupying excessive tokens. |
similarityThreshold | 0.75–0.85 | Financial report data is highly structured. This threshold filters low-relevance document fragments, only recalling content matching core financial report modules. |
recallTopK | Top 3 recall results | Segmented modules for paint and ink financial reports are limited. Excessive recall leads to redundant context. 3 results cover the three core analysis dimensions of revenue, costs, and R&D. |
systemPrompt | Calibrate data units in accordance with paint and ink industry financial report disclosure standards, first clarify target enterprises and report periods before conducting analysis | Must unify disclosure differences across enterprises to prevent conversations from deviating from user needs. |
segmentMaxLen | 1000–1200 characters | Segmented data paragraphs in financial report supplementary documents are lengthy. This segment length retains complete product revenue or cost details, avoiding loss of context after splitting. |
maxContext | 8000 token | Multi-turn dialogue must retain historical context and recalled financial report fragments. 8000 tokens cover complete analysis logic and adapt to window limits of mainstream large models. |
The parameter values provided on this page are common starting points for configuration. Actual values are affected by material form, data volume, and business rules. Specific issues require specific analysis. It is recommended to test on your own samples before finalizing values.
Three Common Mistakes
- Phenomenon: Accessing the
Text Content Extractionplugin in a workflow and configuringChat History: 6 entriesfails to enable multi-turn dialogue. Cause: The workflow node is not bound to global dialogue context storage. Only single-round data is cached within the plugin, and historical conversation information cannot be passed across nodes. - Phenomenon: Calling an interface to obtain multi-turn dialogue results returns
400 Bad Request. Cause: Thehistoryfield is not correctly carried in the request body, or the field format does not match the array structure required by the large model. - Phenomenon: Inconsistent unit comparison results appear in multi-turn dialogue, such as directly adding data in ten thousand yuan and hundred million yuan. Cause: The prompt does not include unit calibration rules for paint and ink financial reports, and fails to unify amount units disclosed by different enterprises.
How to Confirm Correct Configuration
- Initiate a first-round query that includes an enterprise name and report period. Verify whether the system automatically prompts to confirm the disclosure standard, and whether the context retains the first-round information.
- Initiate cross-round follow-up queries, such as first asking about revenue composition before asking about raw material costs. Verify whether the system can associate enterprise and report period information from the previous round.
- Test recall of financial report fragments with different units. Verify whether the system automatically unifies data units before outputting results.
- Call an interface to pass a historical conversation array. Verify whether the returned result is generated based on complete context.
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.