What the data for this category looks like
Marketing-related data for residential development comes primarily from four sources: customer consultation transcriptions from project on-site sales offices, official public documents of unit types and supporting facilities, user question records from online lead generation forms, and version iteration archives of marketing materials. Data update rhythms vary by scenario: on-site consultation records are generated in real time, unit type parameters are only updated when project plans are adjusted, and marketing materials are updated 1-2 times per month based on promotion cycles. Document structures fall into two categories: structured fields such as building number, unit floor area (unit: ㎡), filing unit price (unit: yuan/㎡), and names of surrounding public supporting facilities; unstructured content such as original customer dialogue transcripts and draft marketing promotion copy.
What constraints these characteristics impose on the "multi-turn dialogue and prompt engineering" link
The clarity of structured fields requires multi-turn dialogue to accurately match preset fields, to avoid parameter errors caused by fuzzy extraction. Real-time generated on-site consultation data requires the dialogue context to retain the last 3 or more user questions, ensuring subsequent responses can carry forward previous request details. The version iteration characteristic of marketing materials requires prompt engineering to include version verification logic, ensuring the latest promotional content is called. Unstructured original customer dialogue transcripts require the system to recognize user demands at different stages, from initial unit type inquiries to later price negotiations, and gradually adjust response directions. At the same time, regulated fields such as filing prices must be explicitly prohibited from being modified arbitrarily in the prompt engineering, to ensure response compliance.
How to set the configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
maxContext | Last 6 rounds of dialogue | Residential development marketing conversations typically progress through fewer than 5 rounds of request iteration. Retaining the last 6 rounds covers the complete consultation flow and prevents context loss |
similarityThreshold | 0.75–0.85 | Structured fields such as unit parameters and filing prices need to be matched. A threshold that is too low will introduce irrelevant content, while a threshold that is too high will fail to retrieve relevant supporting information |
contextSplitSize | 800–1200 characters | Marketing material documents are often long texts. The segment length adapts to the context window of large language models, preventing truncation of critical unit type and price information |
maxToken | 8000–12000 | Adapts to the context length of long-text marketing materials and multi-turn dialogues, ensuring complete transmission of user requests and project information |
logRetentionDays | 30 days | Marketing consultations require retaining historical conversations for review. A 30-day retention period covers standard promotional review cycles while controlling storage costs |
promptTemplate | Explicitly specify extraction of fields including building number, unit area, filing unit price, prohibit modification of regulated price information, adjust expression methods to adapt to the context understanding ability of the accessed large language model | Residential development marketing content requires accurate matching of project parameters. The prompt engineering must constrain the response scope to avoid generating non-compliant content |
The parameter values provided on this page are common recommendations for establishing configuration baselines. Actual values are affected by material formats, data volume, and business rules. Specific issues require targeted analysis, and it is recommended to test on your own samples before finalizing settings.
Three common mistakes
- Phenomenon: When the prompt engineering includes field names with spaces (such as "filing unit price"), the system cannot accurately match the corresponding structured data. Cause: The prompt template does not explicitly require retaining spaces, or field names are not wrapped in quotation marks to avoid being split.
- Phenomenon: For applications released via the API channel, subsequent requests cannot carry forward historical conversation content, and the response does not link to previous user questions. Cause: Valid session ID parameters are not included in the API request body, causing each request to be treated as a new conversation.
- Phenomenon: After deployment and container restart, conversation logs older than 7 days cannot be retrieved. Cause: A persistent log directory is not configured in the Docker configuration file, or a reasonable value for
logRetentionDaysis not set, causing logs to be automatically cleaned up or not persistently stored.
How to confirm proper configuration
- Upload a residential project unit type parameter document, enter a query that includes a field name with spaces, and verify whether the response accurately extracts the corresponding field.
- Initiate more than 3 consecutive conversations, check whether the response carries forward details from the previous round of questions, and confirm that the context link is complete.
- Log in to the FastGPT workspace, view the conversation log list, and confirm that logs older than 7 days can still be viewed normally.
- Initiate a multi-turn request via the API channel, include the session ID parameter, and verify whether the response to the second request links to the content of the first request.
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.