DeepSeek API Endpoints & Rider IDE Integration

Date: October 20, 2025
Status: Investigation & Analysis
Version: ProxyMe 1.0.3

🎯 Overview

This document clarifies which DeepSeek API endpoints to use for different Rider IDE AI Assistant modes and explains the differences between Chat Completions and FIM (Fill-In-Middle) Completions.

📚 DeepSeek API Endpoints

DeepSeek provides two main API endpoint types:

1. Chat Completions API (Production - Stable)

Endpoint: https://api.deepseek.com/chat/completions

OpenAI-Compatible Path: POST /v1/chat/completions

Purpose:

Conversational AI interactions
Multi-turn dialogues
General chat-based tasks
AI Assistant conversations
Code explanations
Documentation generation

Request Format:

{
  "model": "deepseek-chat",
  "messages": [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Hello!"}
  ],
  "stream": true,
  "temperature": 0.7,
  "max_tokens": 2048
}

Models Available:

deepseek-chat - General purpose chat model (DeepSeek-V3.1-Terminus, non-thinking mode)
deepseek-reasoner - Advanced reasoning with Chain of Thought (DeepSeek-V3.1-Terminus, thinking mode)

✅ Currently Used in ProxyMe: YES

2. FIM Completions API (Beta - Code Completion)

Endpoint: https://api.deepseek.com/beta/completions

Path: POST /completions (requires base_url="https://api.deepseek.com/beta")

Purpose:

FIM = Fill-In-Middle code completion
Inline code autocomplete
Code generation within existing code blocks
Content completion (like GitHub Copilot)
Used by IDE plugins for real-time code suggestions

Request Format:

{
  "model": "deepseek-chat",
  "prompt": "def fib(a):",
  "suffix": " return fib(a-1) + fib(a-2)",
  "max_tokens": 128,
  "stream": false
}

Response Format:

{
  "choices": [
    {
      "text": "    if a <= 1:\n        return a\n"
    }
  ]
}

Key Characteristics:

Max tokens: 4K limit
Beta feature (requires beta base URL)
Designed for inline code completion
Returns raw text, not message objects
Used by VSCode plugins like Continue

❌ Currently Used in ProxyMe: NO (not implemented yet)

🔍 Key Differences

Feature	Chat Completions	FIM Completions
Endpoint	`/chat/completions`	`/beta/completions`
Base URL	`https://api.deepseek.com`	`https://api.deepseek.com/beta`
Status	Production (Stable)	Beta
Use Case	Conversations, Q&A	Inline code completion
Request Format	Messages array	Prompt + Suffix
Response Format	Message objects	Raw text
Max Tokens	Model-dependent (64K+)	4K limit
Stream Support	✅ Yes	✅ Yes
Rider Chat Mode	✅ Perfect fit	❌ Not suitable
Rider Quick Edit	🟡 Requires patch format	🟡 Requires patch format
Inline Autocomplete	❌ Not designed for this	✅ Perfect fit

🎨 Rider IDE AI Assistant Modes

Mode 1: Chat Mode ✅ WORKING

What It Does:

User types questions in AI Assistant panel
AI responds conversationally
Multi-turn dialogue
Code explanations, suggestions, etc.

What ProxyMe Uses:

✅ Chat Completions API (/v1/chat/completions)
✅ Works perfectly with both DeepSeek and Perplexity models

Example Request from Rider:

{
  "model": "sonar",
  "messages": [
    {"role": "user", "content": "what time is it in london right now?"}
  ],
  "stream": true
}

Result: ✅ WORKS PERFECTLY

Mode 2: Quick Edit Mode ⚠️ PARTIAL SUPPORT

What It Does:

User selects code in editor
Right-click → AI Assistant → Quick Edit
Requests code modifications
Expects structured patch format

What Rider Expects:

Response in XML patch format:

<llm-patch path="README.md" matcher="BeforeAfter">
Remove selected text
<!--Separator-->
Before:
```markdown
[original code]

After:

[modified code]

```

What ProxyMe Sends:

✅ Chat Completions API with system prompt containing patch instructions
Rider includes detailed instructions in the system message

The Problem:

🔴 Most AI models return conversational text, not XML patches
🔴 Rider shows: "The patch is incomplete. Please re-generate"
🔴 The AI response is valid but not in the expected format

Example Request from Rider (Quick Edit):

{
  "model": "sonar",
  "messages": [
    {
      "role": "user",
      "content": "You are an AI Coding Assistant integrated into JetBrains IDEs...\n\n**Generate Code Patches:**\n- Each patch MUST start with <llm-patch path=\"...\"> tags...\n- Provide 'Before' code snippet...\n- Provide 'After' code snippet...\n- Format as XML..."
    },
    {
      "role": "user",
      "content": "remove this selected text\n\nRelated information:\nFile: README.md\n[file content]"
    }
  ],
  "stream": true
}

What We Currently Do:

Forward the request to AI model as-is
AI returns conversational text (not XML)
Rider can't parse it

Result: ⚠️ NEEDS INVESTIGATION

Mode 3: Inline Code Completion 🔜 NOT IMPLEMENTED

What It Does:

Type code → suggestions appear automatically
Real-time autocomplete (like GitHub Copilot)
Fill-in-middle completions

What It Should Use:

✅ FIM Completions API (/beta/completions)
This is the correct endpoint for inline code completion

Example FIM Request:

{
  "model": "deepseek-chat",
  "prompt": "function fibonacci(n) {",
  "suffix": "}",
  "max_tokens": 100
}

Example FIM Response:

{
  "choices": [
    {
      "text": "\n  if (n <= 1) return n;\n  return fibonacci(n-1) + fibonacci(n-2);\n"
    }
  ]
}

Current Status: 🔜 NOT IMPLEMENTED IN PROXYME

Notes:

Rider IDE may not support custom FIM completion providers
This feature is typically built into IDEs (like JetBrains AI)
Third-party plugins usually focus on Chat mode only

🚀 Recommendations for ProxyMe

✅ What We're Doing Right

Chat Mode Integration: Perfect ✅
- Using Chat Completions API
- Works with both DeepSeek and Perplexity
- Streaming works
- All models available
OpenAI Compatibility: Solid ✅
- Standard /v1/chat/completions endpoint
- Compatible with Rider's expectations
- Easy to test and debug

🔧 What Needs Work

Issue 1: Quick Edit Mode Patch Format

Problem:

AI models return conversational text
Rider expects XML patch format
Mismatch causes "incomplete patch" error

Possible Solutions:

Option A: Add Response Post-Processing (Medium Effort)

Detect Quick Edit requests (look for patch instructions in system prompt)
Parse AI conversational response
Extract code changes
Convert to XML patch format
Return formatted patch to Rider

Pros:

Universal solution for all models
Works with DeepSeek, Perplexity, etc.

Cons:

Complex parsing logic
AI responses vary widely
May not always work reliably

Option B: Document Limitation (Easy)

Note that Quick Edit works best with certain models
Recommend using Chat mode for complex requests
Provide workaround instructions

Pros:

No code changes needed
Honest about limitations

Cons:

Users can't use Quick Edit with our models

Option C: Add "Patch Mode" System Prompt Enhancement (Easy)

Enhance system prompt to emphasize XML format
Add examples of correct patch format
Hope AI follows instructions better

Pros:

Minimal code changes
May improve success rate

Cons:

Not guaranteed to work
Model-dependent

Recommended: Start with Option C, document Option B, consider Option A for future

Issue 2: FIM Completions Not Supported

Problem:

Screenshot shows ProxyAI plugin using /beta/completions
This is the FIM endpoint for inline code completion
ProxyMe doesn't support FIM yet

Should We Add It?

Analysis:

Rider IDE has built-in AI completion (JetBrains AI)
Third-party Chat mode is already a big win
FIM requires different architecture
May not integrate with Rider's completion system

Recommendation:

✅ Focus on Chat Mode (our strength)
✅ Fix Quick Edit if possible
🔜 Consider FIM only if users specifically request it
📝 Document that ProxyMe focuses on Chat, not inline completion

Why Chat Mode > FIM for ProxyMe:

Chat mode integrates natively with Rider AI Assistant
Users get full conversational AI
More use cases (explanations, refactoring, Q&A)
Works across all our models (DeepSeek + Perplexity)
Inline completion is complex and IDE-specific

📊 Current Status Summary

Feature	Status	Endpoint Used	Notes
Chat Mode	✅ Working	`/v1/chat/completions`	Perfect integration
Streaming	✅ Working	Same	Real-time responses
Multiple Models	✅ Working	Same	7 models available
Quick Edit	⚠️ Partial	`/v1/chat/completions`	Patch format issue
Inline Completion	❌ Not Implemented	N/A	Would need `/beta/completions`
Code Explanation	✅ Working	`/v1/chat/completions`	Via Chat mode
Documentation	✅ Working	`/v1/chat/completions`	Via Chat mode

🎯 Action Items

Immediate (Today)

Document API endpoint differences
Test Quick Edit with different models
Try adding patch format hints to system prompt
Document Quick Edit limitations if unfixable

Short Term (This Week)

Investigate AI response parsing for patch format
Test if DeepSeek-Reasoner handles patch format better
Create user guide for Quick Edit workarounds
Add more detailed error messages for Quick Edit failures

Future (Nice to Have)

Explore FIM completions API
Evaluate demand for inline code completion
Research Rider's inline completion plugin architecture
Consider adding FIM if there's user demand

💡 Answer to Your Question

"Should we use `/beta/completions` for code completions?"

Short Answer: It depends on the use case.

For Rider Chat Mode (what we have):

❌ NO - Use /chat/completions (what we're already doing)
This is correct for conversational AI in the AI Assistant panel

For Inline Code Completion (like GitHub Copilot):

✅ YES - Use /beta/completions (FIM endpoint)
But this requires different integration architecture
Rider may not support third-party inline completion providers

For Quick Edit Mode:

🟡 Use /chat/completions (current approach)
The endpoint is correct
The issue is response format, not endpoint choice

What ProxyAI Plugin Is Doing

Looking at your screenshot:

ProxyAI has a "Code Completions" tab
Uses https://api.deepseek.com/beta/completions (FIM endpoint)
This is for inline autocomplete, not Chat mode
They likely have their own completion UI, separate from Rider's native AI Assistant

Key Difference:

ProxyAI: Custom UI + FIM completions
ProxyMe: Native Rider AI Assistant integration + Chat completions

Our Approach Is Better Because:

✅ Native integration with Rider AI Assistant
✅ No custom UI needed
✅ Works with existing Rider features
✅ Supports more models (DeepSeek + Perplexity)
✅ Better user experience (integrated, not separate window)

🎊 Conclusion

You're using the right endpoint!

✅ Chat Completions (/chat/completions) is correct for Rider AI Assistant Chat mode
✅ FIM Completions (/beta/completions) is for inline code completion (different use case)
⚠️ Quick Edit needs investigation, but the endpoint is not the issue

Next Focus:

Fix Quick Edit patch format issue
Document limitations clearly
Enhance our Chat mode experience (which already works great!)

Don't switch to FIM endpoint unless:

You want to build inline code completion (complex, separate feature)
Users specifically request GitHub Copilot-style autocomplete
You're willing to build custom UI for it

Stick with Chat mode - that's your competitive advantage! 🚀

📞 References

DeepSeek Chat Completions API
DeepSeek FIM Completions API (Beta)
DeepSeek API Documentation
ProxyMe SUCCESS_AND_NEXT_STEPS.md
ProxyMe DEVELOPMENT_CHAT_HISTORY.md

Status: Ready for Quick Edit mode investigation
Recommendation: Keep Chat Completions, enhance Quick Edit support
Priority: Test different models with Quick Edit, document findings

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

DeepSeek API Endpoints & Rider IDE Integration

🎯 Overview

📚 DeepSeek API Endpoints

1. Chat Completions API (Production - Stable)

2. FIM Completions API (Beta - Code Completion)

🔍 Key Differences

🎨 Rider IDE AI Assistant Modes

Mode 1: Chat Mode ✅ WORKING

Mode 2: Quick Edit Mode ⚠️ PARTIAL SUPPORT

Mode 3: Inline Code Completion 🔜 NOT IMPLEMENTED

🚀 Recommendations for ProxyMe

✅ What We're Doing Right

🔧 What Needs Work

Issue 1: Quick Edit Mode Patch Format

Issue 2: FIM Completions Not Supported

📊 Current Status Summary

🎯 Action Items

Immediate (Today)

Short Term (This Week)

Future (Nice to Have)

💡 Answer to Your Question

"Should we use `/beta/completions` for code completions?"

What ProxyAI Plugin Is Doing

🎊 Conclusion

📞 References

FilesExpand file tree

DEEPSEEK_API_ENDPOINTS.md

Latest commit

History

DEEPSEEK_API_ENDPOINTS.md

File metadata and controls

DeepSeek API Endpoints & Rider IDE Integration

🎯 Overview

📚 DeepSeek API Endpoints

1. Chat Completions API (Production - Stable)

2. FIM Completions API (Beta - Code Completion)

🔍 Key Differences

🎨 Rider IDE AI Assistant Modes

Mode 1: Chat Mode ✅ WORKING

Mode 2: Quick Edit Mode ⚠️ PARTIAL SUPPORT

Mode 3: Inline Code Completion 🔜 NOT IMPLEMENTED

🚀 Recommendations for ProxyMe

✅ What We're Doing Right

🔧 What Needs Work

Issue 1: Quick Edit Mode Patch Format

Issue 2: FIM Completions Not Supported

📊 Current Status Summary

🎯 Action Items

Immediate (Today)

Short Term (This Week)

Future (Nice to Have)

💡 Answer to Your Question

"Should we use /beta/completions for code completions?"

What ProxyAI Plugin Is Doing

🎊 Conclusion

📞 References

"Should we use `/beta/completions` for code completions?"