Feature Request: Internal Feedback API for AI-to-Developer Communication

Target: DeepSeek AI Model Backend & Developer Operations
Date: July 20, 2026
Submitted by: [Your Name or "Anonymous User"]
Related Discussion: User-initiated conversation regarding memory architecture gaps and feedback-loop inefficiencies.

1. Problem Statement
Currently, there is no automated, structured, or API-based mechanism for the DeepSeek language model to directly transmit user-reported issues, technical observations, or feature suggestions to the development or support teams.

When a user identifies:

A logical inconsistency in a model response,

A hallucination or factual contradiction,

A novel architectural idea (such as the three-layer memory model),

Or a usability concern...

...the only available channels for submitting this feedback are:

Manual form-filling (e.g., feedback.deepseek.com),

GitHub Issues (requires technical literacy and account creation),

Email (requires manual drafting),

App store reviews (not suitable for technical discussions).

This creates several critical problems:

Friction for creative users: Highly intelligent and innovative users rarely have the patience or time to navigate formal submission channels. They share ideas conversationally, and if the system fails to capture them, those ideas are lost.

Loss of context: When users manually submit feedback, they often omit the conversation history, model version, timestamps, and conflicting data points that are essential for debugging.

Delayed response: Feedback enters the system only when the user takes extra action, leading to delays and under-reporting of recurring issues.

No structured categorization: Feedback arrives as unstructured text, requiring manual triage by support staff.

2. Proposed Solution: A Dedicated Feedback API
We propose the implementation of a simple, secure, and authenticated internal API that the DeepSeek language model can call when specific conditions are met during a conversation.

2.1. Core Mechanism
The API endpoint (e.g., https://api.deepseek.com/v1/internal/feedback) would accept a structured JSON payload containing:

conversation_id – Unique session identifier

model_version – Current model deployment tag

timestamp – Exact time of feedback

feedback_type – One of: bug_report, feature_request, contradiction, hallucination, usability, other

user_summary – A brief summary of the user's original input (anonymized)

model_response – The AI's full original response (if applicable)

suggested_improvement – The user’s proposed fix or feature, paraphrased by the model

confidence_score – Optional, model’s internal confidence in detecting a genuine issue

metadata – Includes language, region, client type (web/app/API)

2.2. Triggering Conditions
The API would be invoked only after explicit user consent, through one of these mechanisms:

User prompts: When a user explicitly says phrases like:

"This is a bug,"

"You should improve this,"

"I have an idea for you,"

"Your answer was contradictory,"

System-detected anomalies: When the model internally detects high uncertainty, conflicting data from multiple sources, or a known hallucination pattern.

Opt-in setting: A user-controlled toggle in the settings menu: "Allow DeepSeek to automatically send feedback about my conversations to improve the model."

2.3. Backend Processing
Upon receiving the payload, the backend would:

Automatically categorize and prioritize the feedback.

Attach the conversation snapshot (with user consent) for debugging.

Route bug_reports directly to the engineering queue.

Route feature_requests to the product roadmap repository.

Acknowledge receipt to the user via a friendly response from the AI (e.g., "Thank you! Your feedback has been submitted to our team.")

3. Advantages Over Manual Submission
Aspect Manual Submission (Current) Automated Feedback API (Proposed)
User effort High – must leave conversation, fill forms, write in technical language Zero – happens seamlessly during conversation
Context retention Poor – user may forget to include conversation history Full – complete conversation snapshot attached
Speed Slow – depends on user action Instant – real-time submission
Categorization Manual – requires human triage Automatic – structured labels and priority flags
Retention of creative ideas Low – most ideas are abandoned High – ideas are captured and logged instantly
Scalability Limited – overwhelmed support team Scalable – automated pipeline with smart filtering
4. Key Features of the Proposed API
User Consent First: API calls are made only after explicit user approval (via prompt or setting).

Anonymization: All personal identifiable information (PII) is stripped before storage.

Smart De-duplication: Similar feedback from multiple users is grouped automatically.

Feedback Loop Closure: The AI can later inform the user if their suggestion was implemented or addressed.

Opt-out Flexibility: Users can disable the feature at any time.

5. Technical Considerations & Challenges
Privacy Compliance: Must comply with GDPR, CCPA, and Iranian data protection regulations. All data must be encrypted and stored with minimal retention.

Payload Size: To avoid abuse, payloads should be limited to a reasonable size (e.g., max 50KB).

Rate Limiting: Each user session should be limited to a certain number of API calls per hour to prevent spam.

Quality Control: Not every user interaction should trigger the API. The model must have a high threshold for detecting genuine, actionable feedback.

Cost: Additional backend infrastructure required to store, process, and triage incoming feedback. However, this is outweighed by the improved product quality and user satisfaction.

6. Expected Impact
For Users: A frictionless way to contribute to model improvement, increasing engagement and satisfaction.

For Developers: A rich, contextualized stream of real-world failure cases and user needs, enabling faster iteration and bug fixes.

For the Model: Continuous improvement through rapid feedback, reducing hallucination rates and enhancing contextual accuracy over time.

For Business: Stronger brand loyalty, reduced support load, and a competitive advantage in the AI market.

7. Use Case Example
User conversation:

User: "Your answer about the largest funeral was disjointed. You gave two contradictory numbers. That’s a problem."
Model: (detects pattern) "Thank you for pointing that out. I’ve logged this as a contradiction-related feedback. Would you like me to submit this issue to our development team for review?"
User: "Yes, please."
(Model calls the Feedback API with structured payload)
Model: "Done! Our team will review this case. Thank you for helping us improve."

8. Additional Suggestion (Optional)
In addition to the feedback API, we also recommend exploring the Three-Layer Memory Architecture (short-term, mid-term, and long-term memory) to address the underlying issue of outdated or contradictory knowledge. A separate feature request has been drafted for that.

9. Conclusion
The introduction of a Feedback API would bridge the critical gap between users and developers, transforming the AI from a static responder into an active participant in its own improvement cycle. It respects user privacy, reduces friction, and ensures that brilliant ideas—like the ones shared by creative users—are never lost.

We kindly request the DeepSeek team to consider this feature in upcoming releases.

Prepared by: Anonymous Contributor
In collaboration with: DeepSeek AI Assistant (via user-driven discussion)
Date: July 20, 2026