# River API Documentation

> Train and serve large language models (LLMs) and agents with River. Learn supervised fine-tuning, reinforcement learning, and the objectives, data formats, and evaluation methods behind effective experiments.

## Getting started

- [Overview](https://docs.river.ai/index.md): Train and serve large language models (LLMs) and agents with River. Learn SFT, reinforcement learning, and how to evaluate your experiments.
- [Quickstart](https://docs.river.ai/quickstart.md): Install the client, connect to River, and generate your first response before starting a training run.
- [How the API works](https://docs.river.ai/guides/api-basics.md): Create a model instance, sample from its weights, and train it with one call.
- [Models and access](https://docs.river.ai/guides/models.md): Choose an available model and connect to River Cloud or your own deployment.

## Inference

- [Models and pricing](https://docs.river.ai/guides/inference-models.md): Available model IDs, token prices, capabilities, and context limits.
- [Quickstart](https://docs.river.ai/guides/inference.md): Call hosted models with the OpenAI SDK or HTTP, including streaming responses.
- [API controls](https://docs.river.ai/guides/inference-controls.md): Model-specific reasoning, sampling, structured output, and Responses compatibility.
- [Images](https://docs.river.ai/guides/inference-images.md): Send images with text to a hosted model.
- [Tools](https://docs.river.ai/guides/inference-tools.md): Define functions and return tool results to the model.
- [Limits and support](https://docs.river.ai/guides/inference-limits.md): Request limits, timeouts, retries, service status, and data handling.

## Supervised fine-tuning

- [Learning from examples](https://docs.river.ai/guides/sft-concepts.md): Understand demonstrations, next-token loss, training masks, and evaluation.
- [Your first SFT run](https://docs.river.ai/guides/sft.md): Train a small supervised task, sample the result, and save your weights.
- [Choosing an adapter](https://docs.river.ai/guides/lora.md): Choose LoRA capacity and the parts of the model to adapt.

## Reinforcement learning

- [Learning from rewards](https://docs.river.ai/guides/rl-concepts.md): Work through rewards, group-relative advantages, and a policy update.
- [Your first RL run](https://docs.river.ai/guides/rl-sync.md): Train a synchronous math recipe with the River RL library.
- [Tools and environments](https://docs.river.ai/guides/rl-tools.md): Extend a rollout into a multi-turn interaction with tools and stateful environments.
- [Image understanding](https://docs.river.ai/guides/image-understanding.md): Include visual observations in agent rollouts and manage image handles across training and evaluation.
- [Asynchronous training](https://docs.river.ai/guides/rl-async.md): Overlap sampling and training with bounded policy staleness.
- [Policy versions](https://docs.river.ai/guides/rl-policies.md): Understand trajectory pinning, masking, group completion, and KV reuse.
- [Evaluation and recovery](https://docs.river.ai/guides/rl-checkpoints.md): Evaluate held-out tasks and recover an interrupted training run.
- [Build your own RL system](https://docs.river.ai/guides/rl-primitives.md): Build custom algorithms with explicit sampling, token-level credit, gradient accumulation, and policy-version control.

## Distillation

- [Distillation overview](https://docs.river.ai/guides/distillation.md): Choose a teacher–student recipe and a path through KL, self-distillation, and specialist teachers.
- [Distill a larger model](https://docs.river.ai/guides/distillation-basics.md): Train a smaller student with sampled reverse-KL feedback from a stronger teacher.
- [On-policy self-distillation](https://docs.river.ai/guides/distillation-self.md): Give the teacher a checked solution while the student learns to answer without it.
- [Self-distillation with River OPSD](https://docs.river.ai/guides/distillation-opsd.md): Align student and teacher contexts for River’s native full-vocabulary forward-KL loss.
- [Multi-teacher distillation](https://docs.river.ai/guides/distillation-multi-teacher.md): Route tasks to specialist teachers and combine their capabilities in one student.

## Working with River

- [Sessions and requests](https://docs.river.ai/guides/requests.md): Understand model lifetimes, blocking calls, and submit-and-poll requests.
- [Metrics and charts](https://docs.river.ai/guides/metrics.md): Report training metrics, choose chart axes, and compare runs in River Console.
- [Save and resume weights](https://docs.river.ai/guides/checkpoints.md): Save training or inference weights and restore them in another session.
- [Deploy and serve](https://docs.river.ai/guides/deployments.md): Deploy a checkpoint and stream responses through an OpenAI-compatible endpoint.
- [Throughput and the Console](https://docs.river.ai/guides/operations.md): Measure batch throughput, inspect runs, and download trained adapters.

## Reference

- [Loss functions](https://docs.river.ai/guides/losses.md): Built-in training objectives, their input fields, and configuration.
- [Python API](https://docs.river.ai/python-api.md): Reference for river-client: core sampling and training primitives, plus the multi-turn RL library.