adrida/tracer

Health v2 · maintenance only

Flags: young

Adoption not part of the score

1030 stars · 76 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

TRACER is a Python library that learns a routing policy from LLM classification traces, fitting a lightweight traditional ML surrogate to handle easy inputs and deferring only uncertain ones back to the LLM. It claims 90%+ of classification calls routed to non-LLM models with formal parity guarantees and a self-improving policy as deferred calls generate new training traces.

Use cases

• reduce LLM API costs for classification pipelines
• route easy classification inputs to a cheap ML model
• distill an LLM classifier into a traditional ML surrogate
• cut inference spend on high-volume text classification
• build a self-improving routing layer over an LLM classifier
• guarantee agreement between surrogate and teacher LLM outputs

When to choose

• you run high-volume LLM classification and want to cut API costs
• most of your classification traffic is repetitive and easy
• you need formal guarantees that a cheaper model matches your LLM's outputs
• you want the routing policy to improve automatically over time

When to avoid

• your inputs are diverse and genuinely require LLM reasoning on most calls
• you need generative outputs rather than classification labels
• you cannot produce labeled traces from your LLM pipeline
• your classification task changes frequently, invalidating learned boundaries

Facets

library · maturity active

machine-learning llm-inference rag machine-learning large-language-models developer-tools python cli llm-cost-optimization model-routing surrogate-model classification distillation cost-reduction

Member repositories

For agents

markdown · JSON · MCP: product_card(name="adrida/tracer")