Prototype agentic guideline search with LiteRT-LM tool calling

[nmrenyi]

Why

The current app uses a manual retrieval on/off switch and injects retrieved context before generation. With Gemma 4 + LiteRT-LM, the app can potentially move to a tool-driven workflow where the model decides when to call retrieval.

This should be treated as a controlled migration, not an unmeasured refactor.

Scope

Prototype a first tool-driven workflow for guideline search, for example:

• search_guidelines(query, top_k) -> returns retrieved chunks / metadata

Keep the first tool set narrow and read-only. Do not let the model autonomously trigger UI actions like opening PDFs.

Implementation direction

• Use LiteRT-LM tool support in the Android runtime
• Start behind a feature flag
• Keep the current manual retrieval path as fallback
• Reuse the existing offline retrieval stack / SQLite vector store

Dependencies

• Follows the exploration in Explore Gemma 4 tool use / function calling capabilities #35
• Should be evaluated only after current benchmark and retrieval baselines are in place

Evaluation requirements

Compare against the current workflow on the same benchmark line:

• retrieval off
• current manual retrieval-on flow
• tool-triggered retrieval flow

Acceptance criteria

• There is a minimal tool registry in the Android pipeline.
• The first tool is retrieval, not arbitrary app actions.
• Behavior is benchmarked against the current manual search workflow before any default switch.
• Failure modes are documented, especially around safety and unsupported questions.

[nmrenyi]

Priority: 4. Depends on #35 for tool-calling exploration and should come after #43, #42, and #41 so the current baseline and retrieval quality are measured first.