What the data for this category looks like
Data for heating marketing comes primarily from user behavior logs, payment bills, maintenance work orders, and heating usage metering data from heating supply systems. Update frequencies cover real-time (user behavior, maintenance status), hourly (metering summaries), and daily (bill statistics). Data is stored as structured tables, with core fields including unique user identifier, heating usage (unit: gigajoules/cubic meter), payment status, last maintenance request time, and historical marketing outreach records. Fields are explicitly bound to corresponding business scenarios, with no redundant mixed fields.
What constraints these characteristics impose on multi-turn dialogue and prompt engineering
The multi-dimensional structured nature of heating data requires prompts for multi-turn dialogue to clearly distinguish the meaning of different business fields, avoiding confusion between heating usage and payment status. The real-time and hourly update rhythm requires the dialogue's knowledge base retrieval to support dynamic refreshing, rather than relying on static cached content. Historical marketing outreach records must be included in the context window to avoid repeatedly pushing the same marketing content. Additionally, since fields are bound to specific business scenarios, prompts must limit retrieval to the current user's exclusive data to prevent cross-user information leakage or confusion.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
stream_mode | Enabled | Heating marketing conversations need real-time feedback on users' inquiries about heating usage and payment progress. Streaming output reduces user perception of waiting time |
maxContext | Previous 10 context entries | Heating data includes multi-dimensional user behavior. Retaining the first 10 entries fully covers users' historical inquiries and heating usage information, avoiding truncation of critical context |
| Retrieval count | Top 8 entries | The heating marketing knowledge base covers core scenarios such as usage inquiries, payments, and maintenance requests. The top 8 entries meet the precise retrieval needs of most user inquiries |
| Similarity threshold | 0.75–0.85 | Heating-related inquiries require precise matching of users' specific usage scenarios. A threshold that is too low will introduce irrelevant content, while a threshold that is too high will miss valid results for relevant scenarios |
custom_quick_questions | Configure preset questions such as "Query heating usage", "Query payment records", "Schedule heating maintenance" | Support users to trigger corresponding processes with one click, matching the marketing needs of rapid customer acquisition |
PARSE_FILE_TIMEOUT_SECONDS | 300 seconds | The heating marketing knowledge base may contain batch user behavior reports and metering data. A longer timeout ensures complete parsing of large-volume files |
The parameter values provided on this page are common recommended starting points for configuration. Actual values are affected by material form, data volume, and business rules. Specific issues require specific analysis, and it is recommended to test on your own samples before finalizing.
Three common mistakes
- Issue: After adjusting the
maxContextparameter, context memory does not take effect, and retrieval results cover historical content from non-current sessions. Cause: The context retrieval scope is not limited to the chat records of the current session, and the retrieval threshold is set too low, introducing irrelevant content from the external knowledge base. - Issue: No streaming output appears after enabling
stream_mode, and conversations still return full paragraphs at once. Cause: For version v4.8.10, the streaming output switch was not manually checked in the advanced settings of the dialogue node, or the basic configuration for streaming transmission was not enabled during deployment. The default setting may not enable this function. - Issue: Clicking a custom quick question does not trigger the corresponding process. Cause: The trigger keywords for the quick question do not match the knowledge base's retrieval rules, or the jump node for question triggering was not configured.
How to confirm configuration is complete
- Open the dialogue page, enter a question related to heating usage, and observe whether returned content is output in segmented streaming format to confirm the
stream_modeconfiguration takes effect. - Initiate 3 consecutive inquiries related to heating payments and usage, check whether the chat record area fully retains the key information from previous conversations to confirm the
maxContextparameter configuration meets expectations. - Click the preset custom quick questions, confirm that the dialog box automatically fills in the corresponding question and triggers knowledge base retrieval to verify the
custom_quick_questionsconfiguration is correct. - For version v4.8.10, upload a heating user behavior report, wait for the parsing to complete, and initiate related inquiries to confirm that the knowledge base can retrieve content from the report, verifying the
PARSE_FILE_TIMEOUT_SECONDSconfiguration is sufficient.
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.