Similarityoung opened a new issue, #937:
URL: https://github.com/apache/dubbo-go-pixiu/issues/937

   ### ✅ Verification Checklist
   
   - [x] 🔍 I have searched the [existing 
issues](https://github.com/apache/dubbo-go-pixiu/issues) and confirmed this is 
not a duplicate
   
   ### ✨ Proposal Summary
   
   ## Background
   
   Pixiu supports exposing backend API services as MCP tools. In real-world 
scenarios, a gateway may manage many backend APIs, which may result in a large 
number of MCP tools.
   
   If all tools are directly exposed to the LLM, it may cause tool overload, 
increase context size, and make tool selection less accurate or less 
controllable.
   
   Therefore, I would like to propose exploring an intelligent MCP tool routing 
mechanism in Pixiu.
   
   ## Proposal
   
   Pixiu could introduce a routing layer to select a smaller and more suitable 
set of MCP tools for the current request or session, instead of exposing all 
available tools directly.
   
   A possible flow:
   
   ```text
   Backend API Services
           ↓
   Pixiu wraps APIs as MCP Tools
           ↓
   MCP Tool Router
           ↓
   Expose selected tools to LLM / Agent
   ```
   
   ### 🛠️ Implementation Approach
   
   This issue does not propose a fixed implementation. Some possible directions 
include:
   Method | Pros | Cons | Suitable Stage
   -- | -- | -- | --
   Rule-based routing | Simple and controllable | Poor expression coverage | MVP
   Workflow bundle | Stable and complete tool chain | Requires manual 
maintenance | Recommended
   Schema matching | Accurate and explainable | Depends on entity extraction | 
Recommended
   Capability mapping | Decouples routing from concrete tools | Requires 
capability model design | Mid/Late stage
   Progressive tool disclosure | Exposes tools gradually like skills | Requires 
session state management | Recommended
   LLM router | Strong semantic understanding | Cost and stability concerns | 
Mid/Late stage
   Embedding-based routing | Good generalization | Not suitable for security 
decisions | Mid/Late stage
   Historical feedback | Improves over time | Weak cold start | Late stage
   
   </div></div>
   <p data-start="1856" data-end="2103" data-imt-p="1">These methods can also 
be combined. For example, Pixiu may first apply policy-based filtering, then 
use workflow bundles, schema matching, capability mapping, progressive 
disclosure, embedding search, or an LLM router to select the final tool set.</p>
   
   ### 📚 Additional Context
   
   ### Expected Benefits
   
   Reduce MCP tool overload for LLMs.
   Improve tool selection accuracy.
   Make MCP tool exposure more controllable.
   Provide a foundation for API-to-MCP governance in Pixiu.
   Enable future policy control, audit, and observability for MCP tool calls.
   
   ### Notes
   
   This issue is only a proposal and direction for discussion. The final design 
and implementation can be further explored by contributors.


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to