[FEATURE]: Enable Dynamic Model Selection (Gemini 3 & Gemma Support)

[TheAnshulPrakash]

Feature and its Use Cases

Description

Currently, DocPilot uses a hardcoded model for processing medical documentation (summarization and prescription generation). While functional, this limits users from leveraging the specialized capabilities of the broader Google AI ecosystem.

With the release of the Gemini 3 series and the high-performance Gemma open-weight models, the application would benefit significantly from allowing users to choose a model that fits their specific latency and reasoning requirements.

Problem

The current implementation: Does not allow switching between Gemini 3 Flash (speed) and Gemini 3 Pro (complex reasoning).

Lacks support for Gemma models, which are highly efficient for straightforward summarization tasks.

Forces a "one size fits all" approach, which may not be ideal for varied clinical workflows.

Proposed Solution

Add a configuration option in the UI (e.g., in a "Model Settings" or "AI Configuration" section) that allows the user to select their preferred model.

Specifically:

• UI Component: A dropdown or selection menu to choose between:

  * Gemma 3 (Open Weights) (Default)
  
  * Gemini 3 Pro
  
  * Gemini 2.5/3 Flash
  

• Dynamic API Calls: Update the AI service layer to pass the selected model_id to the Gemini API client during inference.

Benefits

Flexibility: Users can prioritize speed (Flash/Gemma) or accuracy (Pro) depending on the complexity of the medical case.

Cost/Resource Optimization: Gemma models provide a generous and efficient alternative for standard tasks.

Additional Context (After updates)

• Response by Gemma-3-27B

Code of Conduct

• I have joined the Discord server and will post updates there
• I have searched existing issues to avoid duplicates

[SharkyBytes]

Makes sense, but we should prioritize local models over external APIs like Gemini to stay aligned with open-source principles
self-hosted/API key access can be kept as an optional fallback.

[TheAnshulPrakash]

Hey @SharkyBytes ,
thanks for replying

Yeah local models definitely make more sense from an open-source POV. What I tried initially was just building on top of the existing app and adding models with better free-tier limits but the current structure honestly isn’t designed to support local LLMs properly without reworking a lot of core parts.

Because of that I’ve been thinking more in terms of redesigning DocPilot itself for GSoC. Not just model selection, but making it more context-aware, like factoring in the type of clinic/medicine domain it’s being used in.

On the implementation side, local LLMs do seem doable:

• packaging models (llama.cpp) and running them via Flutter across devices
• using faster-whisper for on-device speech-to-text
• using something like Gemma 3 4B for summarization

One thing I’m not fully sure about is the expected scope though,
are we targeting strictly mobile/edge devices, or can we assume users might have access to more capable systems like laptops? That would affect model choices quite a bit.

[SharkyBytes]

yeah we targeting strictly mobile edge devices for now

[TheAnshulPrakash]

ok,
Thanks for replying