This document is intended to explain the capability boundaries, system limitations, and known risks of AI-Slides in business use, helping business teams establish reasonable expectations.
1. Key Execution Limits
|
Limitation Item |
Default Value |
Business Impact |
|---|---|---|
|
Maximum background research steps |
30 steps |
Research will stop after this limit is reached, and generation will continue based on existing information, which may result in incomplete information |
|
Background research timeout |
15 minutes |
After timeout, the system will not continue waiting and will directly use the collected content to continue generating the outline |
|
Number of results returned per search |
5 |
Search coverage is limited, and long-tail information may not be included |
|
Maximum outline pages |
30 pages |
Requests exceeding 30 pages will be truncated |
|
Image generation concurrency |
3 |
In scenarios with many images, tasks will be queued, extending the overall completion time |
|
Uploaded image size |
10 MB/image |
Files exceeding the limit cannot be uploaded |
|
Supported image formats |
jpeg/png/webp/gif |
Other formats cannot be uploaded |
|
Single model request timeout |
120 seconds |
Timed-out requests will be retried; if they still fail, an error will be reported |
|
Model request retry count |
2 times |
After consecutive failures, the task will be interrupted or partially fail |
|
Gateway single request duration |
600 seconds |
Extremely long requests may be terminated by the gateway |
2. Business Differences Between Two Task Modes
1) Conversational Outline Generation
-
Suitable for scenarios where users want to "chat and revise as they go"
-
Supports interruption, confirmation, modification, and regeneration
-
Session state can be continued, making it suitable for multi-round refinement
-
Routing is autonomously decided by the model: it may generate the outline directly, or it may conduct background research first
2) Batch Slide Generation
-
Suitable for "one-time bulk draft generation" scenarios
-
Executed asynchronously in the background; tasks usually continue even after the frontend disconnects
-
Supports reconnecting to continue viewing progress
-
Supports retrying only failed pages without rerunning the entire set
Notes
-
Conversational mode is more flexible, but the path is not fully controllable
-
Batch mode is more stable, but "partial page failures" may still occur
3. Boundaries of Interruption and Manual Confirmation
The system will require manual confirmation at key nodes. Common scenarios include:
-
Confirmation after outline generation (confirm/modify/regenerate)
-
Template selection confirmation
-
When regenerating without a clear direction, the system will provide suggestions for selection
Please note: AI-Slides is not a "fully automated one-click end-to-end" system. Manual participation is required at key steps. If "regenerate" is selected multiple times without providing a clear direction, the process may enter repeated waiting.
4. Image Generation Boundaries
-
The maximum image concurrency per task is 3
-
Failed image generation is not automatically retried by default
-
Image failures do not block main text generation; the main text can still be completed first
Therefore, the final result may show a situation where "the copy is complete but some images are missing," requiring manual image supplementation or single-page retry.
5. Permission and Data Isolation Boundaries
Permission Boundaries
-
Template management (create/modify/delete) belongs to super admin permissions
-
Regular users cannot operate super admin APIs
Data Isolation Boundaries
-
Users can only access their own data by default
-
User identity is determined by the system authentication result, and frontend self-assignment is not accepted
This means that cross-account data viewing or operations will be blocked. When permissions are insufficient, the operation will fail directly, and there will be no "degraded pass-through."
6. Boundaries of External Tools and Dependencies
Background research depends on external tools (search, knowledge base, MCP, etc.):
-
When a certain tool is unavailable, the system may skip that tool and continue execution
-
When research times out or is interrupted, output will continue based on the "partially obtained information"
In such cases, the task may not necessarily fail, but the completeness and depth of the output may decline, and results for the same question may also fluctuate at different times.
7. Multilingual Boundaries
-
The system supports Chinese/English/Japanese/Korean interfaces and error messages
-
The language of AI-generated content is mainly determined by task settings and prompts
Please note that the interface language can be switched, but this does not mean the AI content will necessarily strictly match the same language. Abnormal output language usually requires adjusting prompts or task parameters.
8. Model and Result Stability Boundaries
1) Risks of Large Page Counts / Complex Templates
-
When AI-Slides generates a full set of content in a single run, it places high demands on the model's context and output capabilities
-
The more pages there are and the more complex the template is, the more likely truncation or partial page failures will occur
2) Path Uncertainty
-
Research and generation paths are dynamically decided by the model
-
The same input may produce different organizational structures and expressions
3) Partial Result Mechanism
-
When research times out, image generation fails, or individual pages fail to generate, the system may return partially usable results first
Business Recommendations
-
For important scenarios, prioritize batch generation in segments (for example, by chapter)
-
Reserve time for manual review and supplementation of key pages
9. Verified Model Scope
The models currently validated for functionality include:
-
OpenAI: GPT-4.1, GPT-5.2, GPT-5.4
-
Anthropic/AWS: Claude (Bedrock)
-
Google: Gemini 2.5, Gemini 3
-
Qwen: qwen3.6-plus-no-think, qwen3.6-flash, qwen3.6-max-preview