Skip to content

Evidence Map

Inspection scope

Inspected the project manifest, README, Vite configuration, main React shell, voice-agent hook, LLM runtime, TTS facade, model/tool libraries, public workers, and parser tests. Generated files, dependency folders, and downloaded model weights were excluded.

Key evidence

Area Evidence
Product README.md, src/App.tsx
Orchestration src/hooks/use-voice-agent.ts
LLM boundary src/lib/llm-runtime.ts, src/lib/llm/
STT/VAD public/stt-worker-esm.js, public/vad-processor.js, src/lib/stt-models.ts
TTS src/hooks/use-tts.ts, src/lib/tts.ts, src/lib/tts-providers/
Tools src/lib/tools/registry.ts, src/lib/tools/execute.ts
Configuration package.json, vite.config.ts, vercel.json, tsconfig*.json
Quality src/lib/llm/parsers/test-parsers.ts, test-tools.ts

Commands used

  • git clone https://github.com/Hari31416/local-voice-chat.git repos/local-voice-chat
  • Repository file and manifest inspection
  • pnpm install and pnpm build are the intended validation commands; model inference was not run in this environment.

Confidence assessment

Subsystem Confidence Reason
UI composition High Main shell and studio components are present
Voice orchestration High useVoiceAgent exposes the complete lifecycle
LLM routing High Adapter registry and tool loop are explicit
STT worker internals Medium Worker is present but model execution depends on browser assets
Persistence Medium Preference/cache helpers are visible; browser storage was not exercised
Deployment Medium Static build and Vercel headers are documented
Security posture Medium Client-only boundary is clear, but browser permission behavior is runtime-dependent

Fact classification

  • Confirmed: no application API server appears in the manifest or source layout.
  • Confirmed: four LLM engine families are represented by adapters.
  • Confirmed: microphone processing and STT run outside the React render thread.
  • Inferred: browser caches make repeated model startup cheaper; exact eviction behavior depends on the provider and browser.
  • Unknown: production cache hit rates, device compatibility across all model variants, and real-world memory pressure.