AWS Bedrock Provider on Bifrost

1) Chat Completions

Primary path for Claude, Nova, Mistral, Llama, Cohere, and Titan chat models. Bifrost maps OpenAI parameters to Bedrock converse. Upstream: Converse API. Images must be base64 or data URI (remote URLs are not supported).

Silently dropped: frequency_penalty, presence_penalty, logit_bias, logprobs, seed, parallel_tool_calls.

curl -X POST http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "bedrock/anthropic.claude-3-5-sonnet-20241022-v2:0",
    "messages": [{"role": "user", "content": "Hello"}],
    "max_completion_tokens": 1024,
    "temperature": 0.7
  }'

2) Responses API

Uses the same underlying converse endpoint with conversion between OpenAI Responses format and Bedrock messages. Gateway: POST /v1/responses.

• String input wraps as a user message; arrays convert to messages
• stopReason maps to response status (completed / incomplete)
• Streaming: response.created → content_part.delta → output_item.done

curl -X POST http://localhost:8080/v1/responses \
  -H "Content-Type: application/json" \
  -d '{
    "model": "bedrock/anthropic.claude-3-5-sonnet-20241022-v2:0",
    "input": "Hello, how are you?",
    "instructions": "You are a helpful assistant",
    "max_output_tokens": 1024
  }'

3) Text Completions

Legacy completions via Bedrock invoke. Streaming is not supported. Only Claude (Anthropic) and Mistral models are supported.

• Claude: prompt wrapped as \n\nHuman: {prompt}\n\nAssistant:; max_tokens → max_tokens_to_sample
• Mistral: standard mapping; multiple outputs in choices[]
• Response: choices[].text; stopReason → finish_reason

curl -X POST http://localhost:8080/v1/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "bedrock/anthropic.claude-3-haiku-20240307-v1:0",
    "prompt": "The capital of France is",
    "max_tokens": 64
  }'

4) Embeddings

Titan and Cohere embedding models via invoke. Gateway: POST /v1/embeddings. Non-streaming only.

curl -X POST http://localhost:8080/v1/embeddings \
  -H "Content-Type: application/json" \
  -d '{
    "model": "bedrock/amazon.titan-embed-text-v2:0",
    "input": "Hello world"
  }'

5) Image Generation

Titan Image Generator v1/v2, Nova Canvas, and Stability AI models (flat JSON when model ID contains stability.). Upstream: invoke. Non-streaming.

curl -X POST http://localhost:8080/v1/images/generations \
  -H "Content-Type: application/json" \
  -d '{
    "model": "bedrock/amazon.nova-canvas-v1:0",
    "prompt": "A futuristic cityscape with a flying car",
    "size": "1024x1024",
    "n": 1
  }'

6) Image Edit

multipart/form-data (not JSON). Titan / Nova Canvas task types:

Stability AI edit models (model ID contains stability.) support inpaint, outpaint, remove-background, upscale, and more. See Image Edit.

curl -X POST http://localhost:8080/v1/images/edits \
  -F "model=bedrock/amazon.nova-canvas-v1:0" \
  -F "type=inpainting" \
  -F "image[]=@photo.png;type=image/png" \
  -F "mask=@mask.png;type=image/png" \
  -F "prompt=Replace the masked area with a garden"

7) Image Variation

multipart/form-data. Titan Image Generator and Nova Canvas models. Task type IMAGE_VARIATION. Optional similarityStrength (0.2–1.0) via extra params. Non-streaming.

curl -X POST http://localhost:8080/v1/images/variations \
  -F "model=bedrock/amazon.nova-canvas-v1:0" \
  -F "image[]=@source.png;type=image/png" \
  -F "n=2" \
  -F "size=1024x1024"

8) Files API

S3-backed file operations integrated with Bedrock. Upload uses multipart/form-data. File purpose is always batch; status is always processed.

• POST /v1/files — multipart upload (file required, filename optional)
• GET /v1/files — list (cursor pagination)
• GET /v1/files/{file_id} — retrieve metadata
• DELETE /v1/files/{file_id} — delete
• GET /v1/files/{file_id}/content — download content

curl -X POST http://localhost:8080/v1/files \
  -F "file=@document.pdf" \
  -F "filename=report.pdf"

9) List Models

Lists foundation models via Bedrock control plane (listFoundationModels). No request body. Results respect region, aliases/deployments, and per-key models allowlists; multi-key configs aggregate results.

curl http://localhost:8080/v1/models

• id — model name (with deployment prefix when configured)
• display_name → name
• Pagination via NextPageToken

10) Error handling

Bedrock HTTP errors map to Bifrost error types. Special cases: context cancellation → RequestCancelled; timeouts → ErrProviderRequestTimedOut; streaming errors sent on the stream channel.