What the data for this category looks like
Data for this category primarily comes from game customer service systems, community forums, app store reviews, and game logs. Daily processed documents include player tickets, compensation requests, public opinion daily reports, and standardized response scripts. Fields include player identifiers, problem types, feedback channels, emotional tendencies, and more. Data update cycles are mostly daily or real-time. Single data entries are small in size, but daily output volume is high. Contextual information such as player accounts and historical feedback must be associated.
What constraints do these characteristics impose on the knowledge base retrieval and recall link
Multi-source mixed data input requires the retrieval link to prioritize matching knowledge base entries strongly related to player identity and problem type, avoiding irrelevant public opinion or old tickets from interfering with judgments. Real-time updated documents require high-frequency synchronization configuration to ensure the latest compensation rules and response scripts can be recalled promptly. The lightweight nature of single data entries and high daily output volume require the retrieval link to be lightweight to reduce latency, adapting to the real-time response needs of on-duty operations. At the same time, core time-consuming links such as account verification and compensation rule judgment must associate corresponding specialized document sets to avoid recalling irrelevant content across categories.
How to set the configuration
This configuration applies to FastGPT v4.17.0 and above versions.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Semantic accuracy for matching player tickets and knowledge base response scripts, filters low-relevance content |
limit | Top 5–8 entries | Controls the context length of single-round recall, adapts to the real-time response needs of on-duty operations |
usingReRank | Enabled | Optimizes the ranking of multi-source recall results, prioritizes returning response scripts and ticket cases that match player problems |
datasetSearchUsingExtensionQuery | Enabled | Expands problem semantics, covers ticket scenarios with vague player descriptions, improves recall coverage |
collectionIds | Specify exclusive dataset IDs for player customer service and public opinion daily reports | Isolates cross-category knowledge base content, ensures only documents related to the current business are recalled |
Three common misconfigurations
- Retrieval results include citation links for knowledge base documents, leading to redundant responses. The cause is failure to adjust the retrieval configuration to disable citation output, as the default setting retains document source links.
- When creating doc/docx format file collections, the metadata parameter is not passed correctly, resulting in retrieval results failing to associate chapter title information. The cause is lack of understanding that this parameter is used to mark document structure hierarchy, and chapter-related fields must be passed according to interface requirements.
- AI responses exceed the scope of the knowledge base, generating content not included in the documents. The cause is failure to limit the retrieval scope to the specified dataset, or failure to configure response rules that only call knowledge base content.
How to verify that the configuration is correct
- Check the startup logs of the App/Admin service to confirm that the server environment variables
AIPROXY_API_ENDPOINT(root address of the AI Proxy service, for example http://aiproxy:3000, do not append /v1 or /chat/completions, this is not the single model address of VLLM/Ollama/Xinference, etc.) andAIPROXY_API_TOKEN(AI Proxy administrator token, consistent with the AI Proxy server configuration and cannot be empty, this is not the application API Key, model vendor key, token issued per application, or token count estimated per request) have been correctly configured, with no startup verification errors. These two variables must not be filled in workflow global variables, application API request headers, or ordinary model management pages as a replacement for service deployment configuration. - Submit a test question related to a player ticket, verify that the retrieved results only include documents from the specified customer service and public opinion datasets.
- Upload a test document in doc/docx format, confirm after retrieval that the results include the chapter title information of the main text, verifying that the metadata parameter takes effect.
- Submit an irrelevant question outside the knowledge base, confirm that the AI cannot generate a corresponding response, and only returns content related to the knowledge base.
The configuration values provided are common starting points and should be evaluated against the reader's own sample data.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-10-19.