What the data for this category looks like
The daily output documents of game design and numerical balance work include gameplay design documents, numerical tables, version requirement lists, and balance test records. Data sources mainly include player behavior data, public competitor materials, battle logs, and version feedback. The update frequency is usually once per week or per version iteration. The document structure combines structured numerical fields such as damage coefficients, level thresholds, and unstructured rule descriptions. Some documents also include image content such as battle flowcharts and numerical simulation screenshots.
What constraints these characteristics impose on the knowledge base retrieval and recall link
Documents in this category combine structured numerical content and unstructured rule text, and have a high update frequency. This requires the retrieval system to support incremental synchronization and multi-type content matching. The field accuracy requirements for numerical documents are high; retrieval deviation may directly lead to gameplay imbalance or version rework, so the relevance threshold of retrieval must be strictly controlled. At the same time, image content included in documents must be fully synchronized, otherwise it will affect the readability of Q&A results. In addition, mixed data from multiple sources requires that the retrieval scope be limited to the exclusive document set for game designers, to avoid mixing irrelevant R&D content and interfering with core retrieval results.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Aligns with the semantic matching accuracy required for numerical rule documents, and avoids retrieving irrelevant numerical table entries |
limit | Top 8 results | Balances context length and recall coverage, adapting to the information needs of single-round numerical simulations |
searchMode | Full-text search | Supports matching both numerical fields and rule description text, covering both design documents and numerical tables |
usingReRank | Enabled | Performs secondary ranking on retrieval results, corrects deviations from semantic matching, and improves the ranking priority of highly relevant documents |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands synonymous queries for numerical terms, such as associating "crit multiplier" with "critical hit damage coefficient" |
collectionIds | Specified by the exclusive document set ID for game designers | Limits the retrieval scope, excludes non-designer R&D documents, and reduces invalid retrievals |
Three common configuration errors
- Phenomenon: Images in some documents cannot be displayed in Q&A results, and no
image_urlparameter is included in the returned content fields. Cause: The original external link address of the image was not retained when documents were synchronized to the knowledge base, or the multi-modal content recall configuration was not enabled. - Phenomenon: Retrieval results are empty after calling the API to connect to Confluence. Cause:
collectionIdswas not correctly configured as the designer document set synchronized from Confluence, so the retrieval scope did not cover the target documents. - Phenomenon: Matching results for numerical terms have large deviations, such as "crit rate" being matched to irrelevant event rule documents. Cause:
datasetSearchUsingExtensionQuerywas not enabled, and synonymous expansion queries were not configured for numerical terms.
How to verify the configuration is correct
- Upload a gameplay design document that includes images, launch a query containing the keyword "image", and check whether the returned results include complete
image_urlparameters. - Call the FastGPT v4.17.0 API, pass the specified
collectionIds, and retrieve non-designer documents to confirm that the returned results are empty, verifying that the scope restriction takes effect. - Launch a query for a numerical term, such as "crit multiplier", and check whether the retrieval results include documents related to "critical hit damage coefficient", verifying that the expansion query takes effect.
- Adjust the
similarityparameter to 0.8, retrieve highly relevant numerical table documents, and check whether the sorting of the returned results meets expectations.
The values provided are common starting points and should be measured against the reader's own samples.
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.