AWS Bedrock

Provider bedrock. Adapter over Amazon Bedrock, AWS's managed service that serves models from several vendors (Anthropic Claude, Meta Llama, Amazon Titan/Nova, Mistral, Cohere…) behind a single API and one bill (AWS's). You talk to all of them through the same LLM(...).

pip install "jangada-ai[bedrock]"

• provider=: "bedrock"
• Authentication: standard AWS credentials — AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY and AWS_REGION (or AWS_DEFAULT_REGION). Profiles (AWS_PROFILE), IAM roles and EC2/Lambda metadata also work — the same resolution chain as boto3.
• Model: the Bedrock model ID (or inference profile), e.g. anthropic.claude-3-5-sonnet-20241022-v2:0.

from jangada_ai import LLM

# uses AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY / AWS_REGION from the env
llm = LLM("bedrock", "anthropic.claude-3-5-sonnet-20241022-v2:0")
resp = llm.complete("Summarize {{ topic }} in one sentence.", topic="rafts")
print(resp.text, resp.cost)

Configuration via environment

In your .env (or process variables):

AWS_ACCESS_KEY_ID=AKIA...
AWS_SECRET_ACCESS_KEY=...
AWS_REGION=us-east-1

Region matters: each model ID only exists in some regions, and inference profiles (e.g. us.anthropic.claude-3-5-sonnet-...) enable cross-region routing. boto3-specific arguments (e.g. endpoint_url, region_name) go through extra=.

What it does

• Text and streaming via Bedrock's Converse API (a unified interface across all models).
• Structured output (parse): tool-forcing when the model supports tools; otherwise a JSON instruction + validation with model_validate_json.
• Vision (images=): on vision-capable models (Claude family, Llama vision, Nova).
• Documents (files=): local text extraction before sending.

When to choose Bedrock

When your stack already lives on AWS and you want unified governance, billing and network isolation (VPC), without spreading each vendor's API keys. Combine it as a fallback for direct Claude (Anthropic) to get cross-cloud redundancy. See Providers and Retry and fallback.

What changed in 1.9.0

• Converse fields passed through: extra={"additionalModelRequestFields": ...}, guardrailConfig, promptVariables, requestMetadata, performanceConfig, serviceTier and additionalModelResponseFieldPaths go straight into the call (they used to be silently dropped).
• tool_choice="none" doesn't send the tools (except when the history already has toolUse, which the API requires).
• Streaming: an error mid-stream (throttling, ModelStreamErrorException) becomes a normalized error in astream too, and the stream is closed if you stop consuming it.
• Native tools (Amazon Nova): web_search() becomes nova_grounding and code_execution() becomes nova_code_interpreter (Nova 2). Requires us.* profiles and the bedrock:InvokeTool IAM permission. See Native tools.
• Usage with cache (cache_read_tokens/cache_write_tokens) and a system message given as parts has its text extracted.