# Introduction
Source: https://benchmark.ratel.sh/docs

## What the Ratel benchmark measures

As an agent's tool catalog grows, exposing every tool on every turn bloats the
context window, slows the model down, and degrades tool selection. **Ratel** is the
context engineering platform that fixes this, resolving *only* the tools that matter
for the current turn. This benchmark quantifies that trade-off: for each model and
tool-pool size we report **accuracy, input tokens, and dollar cost** for the Ratel arm
versus a control that exposes the full pool.

<Cards>
  <Card title="BFCL results" href="/docs/bfcl">
    The Berkeley Function-Calling Leaderboard run: accuracy, token, and cost deltas
    across model families (local, open-source, and frontier) and pool sizes.
  </Card>
</Cards>

## How to read the numbers

Each results table compares arms on the same scenarios:

- **control-baseline** — the entire candidate tool pool is exposed to the agent.
- **control-oracle** — only the gold tool is exposed (the accuracy upper bound).
- **ratel** — Ratel discovery + selection: the model sees a handful of targeted tools
  plus the gateway tools, regardless of how large the underlying pool is.

Pool sizes sweep realistic MCP setups (30 → 600 tools). The headline story is that
Ratel's token savings grow with the catalog while accuracy stays at or near baseline.

<Callout title="Built for AI agents too">
  This site publishes [`/llms.txt`](/llms.txt) and [`/llms-full.txt`](/llms-full.txt):
  machine-readable indexes regenerated from this content on every build, so coding
  agents always read the current results.
</Callout>

## Where the data comes from

The results pages are **generated at build time** from
[`ratel-ai/ratel-bench`](https://github.com/ratel-ai/ratel-bench), so the published
site always tracks the source-of-truth report. See the
[BFCL results](/docs/bfcl) for the full breakdown.

---

# BFCL
Source: https://benchmark.ratel.sh/docs/bfcl

Ratel 0.5.0-luna-smoke on the Berkeley Function-Calling Leaderboard.

## Task completion by model and arm

Pooled over simple + multiple splits. Without Ratel = full tool pool in context; With Ratel = search_tools gateway; Oracle = gold tools only (ceiling).

| Model | Arm | Task completion | Tool selection | Recall | Mean total tokens | p50 latency |

| --- | --- | --: | --: | --: | --: | --: |

## Retrieval evaluation (BM25, no LLM)