Aller au contenu principal

ADR-0004: Dual-Path Model Loader Strategy

Status​

Accepted - Q2 2025

Context​

With configuration contracts solidified (ADR-0003), the team needed a deterministic way to toggle between inexpensive stub inference for local development and production-grade large language models (LLMs) for demos and benchmarks. Constraints included:

  • Keep costs near-zero for day-to-day development and CI.
  • Support rapid experimentation with Hugging Face models (e.g., Mistral-7B) without rewriting orchestration code.
  • Provide a graceful path for hardware acceleration (GPU, quantization) while staying compliant with cost and security guardrails from ADR-0005.
  • Ensure the docs portal and evaluation jobs can rely on consistent outputs per environment.

Decision​

Implement a dual-path model loader in ai_core/model_loader.py that:

  1. Defaults to a stub backend returning deterministic, low-latency responses.
  2. Supports a configurable "real" backend (initially mistralai/Mistral-7B-v0.1) using Hugging Face transformers.
  3. Allows optional 4-bit quantization via BitsAndBytesConfig when GPUs are present.
  4. Emits structured metadata (model name, device, quantization flags) to downstream logging for governance.
  5. Respects environment configuration sourced via ADR-0003, including budget gates and secrets.

A single factory function returns a loader object adhering to a small interface (generate, token_usage, close). The choice of backend is controlled by config (environment and tier aware) and can be overridden via environment variables during experiments.

Alternatives Considered​

  • Always-real model
    • Pro: Single path
    • Con: Expensive, slower iteration, brittle CI
  • Separate services for stub vs. real
    • Pro: Strong isolation
    • Con: Operational overhead, duplicated contracts
  • Managed LLM-only approach
    • Pro: Minimal infrastructure
    • Con: Less control over guardrails, harder to test offline

Consequences​

  • Enables local development and CI to run quickly with deterministic outputs.
  • Provides a safe toggle to activate high-fidelity LLMs for demos, benchmarks, or production, aligned with tiered pricing (ADR-0001).
  • Centralizes retry and error handling, making it easy to inject future providers.
  • Requires monitoring to ensure the real backend does not exceed cost ceilings (addressed in ADR-0006 evaluations and ADR-0005 guardrails).

Rollout Plan​

  1. Build the loader abstraction with factory selection and context manager semantics.
  2. Wire config toggles (model.backend, model.quantization) sourced from ADR-0003.
  3. Add pytest coverage for both stub and real paths, including unhappy scenarios (missing weights, OOM).
  4. Integrate telemetry into evaluation harness (ADR-0006) and docs portal demos.
  5. Document usage in /docs-site/docs/github/ai-core.md and update onboarding guides.

Success Metrics​

  • All AI surfaces (CLI, evaluation jobs, docs portal) create loaders via the factory.
  • CI runtime remains under 15 minutes with stub backend.
  • Real backend demos stay within monthly budget thresholds defined in ADR-0005.

References​