Examples

Examples

Examples should be concrete enough to benchmark. Each example needs a baseline, an optimized plan, acceptance criteria, and an evidence bundle.

RAG Deduplication and Cache Layout

Baseline:

1. Retrieve 20 chunks.
2. Send all chunks to a frontier model.
3. Generate answer.
4. Return answer without structured validation.

Migaki:

1. Represent workflow as mIR.
2. Deduplicate overlapping or exact duplicate chunks.
3. Preserve stable instructions as a cacheable prefix.
4. Estimate token and cost deltas.
5. Route lightweight ranking to a cheaper model or mock backend.
6. Use a stronger model for final synthesis.
7. Validate answer against cited chunks.
8. Retry only synthesis if validation fails.
9. Export evidence bundle.

Measures:

• baseline tokens,
• optimized tokens,
• baseline cost,
• optimized cost,
• baseline latency,
• optimized latency,
• validator pass rate,
• plan diff,
• trace and replay artifact.

Code Review Workflow

Baseline:

1. Load repository context.
2. Load changed files.
3. Load style guide.
4. Ask one frontier model for review.
5. Ask the same model to produce final comments.

Migaki:

1. Mark style guide as fixed and cacheable.
2. Mark changed files as non-droppable.
3. Drop unrelated historical conversation.
4. Run static checks deterministically.
5. Route simple lint explanation to a cheaper model or mock backend.
6. Use a stronger model for architectural review.
7. Validate final comments against changed lines.
8. Export evidence bundle.

Measures:

• comment acceptance rate,
• false-positive rate,
• validator pass rate,
• context diff,
• cost delta,
• latency delta.

Structured Output Validation

Baseline:

1. Ask a model for JSON.
2. Parse the response.
3. Retry the entire prompt when parsing fails.

Migaki:

1. Represent schema requirement in mIR.
2. Lower to provider-native structured output when available.
3. Use post-validation fallback when native support is unavailable.
4. Retry only the invalid node.
5. Emit evidence showing the provider capability path.

Measures:

• schema validity,
• retry count,
• cost per valid result,
• backend downgrade warnings.