> ## Documentation Index
> Fetch the complete documentation index at: https://docs.pandaprobe.com/llms.txt
> Use this file to discover all available pages before exploring further.

# AWS Bedrock

> Auto-trace AWS Bedrock Converse and InvokeModel calls

<Note>
  `wrap_bedrock` is currently in **beta**.
</Note>

### Installation

<Tabs>
  <Tab title="pip">
    ```bash theme={null}
    pip install "pandaprobe[bedrock]"
    ```
  </Tab>

  <Tab title="uv">
    ```bash theme={null}
    uv add "pandaprobe[bedrock]"
    ```
  </Tab>
</Tabs>

The `bedrock` extra installs `boto3>=1.34.0`. For async support install `aioboto3` separately — `wrap_bedrock` detects it at runtime and instruments async methods automatically without making it a hard dependency.

### Setup

```python theme={null}
import boto3
from pandaprobe.wrappers import wrap_bedrock

client = wrap_bedrock(
    boto3.client("bedrock-runtime", region_name="us-east-1")
)
```

## Converse API (recommended)

Span name: `"bedrock-converse"`, SpanKind: `LLM`

```python theme={null}
response = client.converse(
    modelId="anthropic.claude-haiku-4-5-20251001-v1:0",
    system=[{"text": "You are a concise assistant."}],
    messages=[
        {"role": "user", "content": [{"text": "Explain recursion in one sentence."}]},
    ],
    inferenceConfig={"temperature": 0.5, "maxTokens": 200},
)
```

The Converse API is provider-agnostic — the same call shape works across Claude, Mistral, Llama, Titan and other Bedrock-hosted foundation models. Prefer Converse over InvokeModel for new integrations.

**What gets traced**

* Input: top-level `system` blocks **hoisted** into the messages list as a `role="system"` entry, followed by the `messages` array. Text-only `content` blocks are flattened into a single string; mixed-block content (images, tool use/results) round-trips as structured JSON.
* Output: assistant `content` text blocks joined together
* Model: `modelId` from the request kwargs
* Token usage (see mapping table below)
* Model parameters: `temperature`, `topP`, `maxTokens`, `stopSequences` from `inferenceConfig`, plus `guardrailConfig`, `additionalModelRequestFields`, `toolConfig`
* `reasoningContent` blocks (when models emit them) are stored in span metadata as `reasoning_summary`

### Streaming

```python theme={null}
response = client.converse_stream(
    modelId="anthropic.claude-haiku-4-5-20251001-v1:0",
    messages=[{"role": "user", "content": [{"text": "Hello!"}]}],
    inferenceConfig={"temperature": 0.5, "maxTokens": 200},
)
for event in response["stream"]:
    delta = event.get("contentBlockDelta", {}).get("delta", {})
    if delta.get("text"):
        print(delta["text"], end="")
```

The wrapper preserves the `{"stream": ..., "ResponseMetadata": ...}` response shape — only the inner iterator is replaced with a tracing-aware reducer. User code accesses `response["stream"]` exactly as before. Time-to-first-token is captured on the first `contentBlockDelta`; final token usage is read from the trailing `metadata` event.

## InvokeModel API (legacy fallback)

Span name: `"bedrock-invoke-model"` (or `"bedrock-invoke-model-stream"`), SpanKind: `LLM`

```python theme={null}
import json

response = client.invoke_model(
    modelId="anthropic.claude-haiku-4-5-20251001-v1:0",
    body=json.dumps({
        "anthropic_version": "bedrock-2023-05-31",
        "max_tokens": 200,
        "messages": [{"role": "user", "content": "Hi"}],
    }),
    contentType="application/json",
    accept="application/json",
)
```

InvokeModel bodies are provider-specific JSON; the wrapper parses the body on a best-effort basis and recognises:

* Anthropic Claude on Bedrock — `{"messages": [...], "system": "..."}`, output `content` blocks, usage as `input_tokens` / `output_tokens`
* Mistral on Bedrock — `{"messages": [...]}`
* Amazon Titan — `{"inputText": "..."}`, output via `results[0].outputText`, usage via `inputTextTokenCount` + `results[0].tokenCount`
* Cohere / Meta Llama — `{"prompt": "..."}` and provider-specific generation fields

Unknown body shapes still produce an LLM span containing the serialised request body as input.

## Async (aioboto3)

`aioboto3` is supported but not required. When `wrap_bedrock` is given an `aioboto3` client (its module path starts with `aioboto3` / `aiobotocore`, or its methods are coroutine functions), the wrapper installs async-shaped patches for `converse`, `converse_stream`, `invoke_model`, and `invoke_model_with_response_stream`.

```python theme={null}
import aioboto3
from pandaprobe.wrappers import wrap_bedrock

session = aioboto3.Session()
async with session.client("bedrock-runtime", region_name="us-east-1") as client:
    wrap_bedrock(client)
    response = await client.converse(...)
```

## Token usage mapping

| Bedrock Field                                 | PandaProbe Field        |
| --------------------------------------------- | ----------------------- |
| `usage.inputTokens` (Converse)                | `prompt_tokens`         |
| `usage.outputTokens` (Converse)               | `completion_tokens`     |
| `usage.totalTokens` (Converse)                | `total_tokens`          |
| `usage.cacheReadInputTokens`                  | `cache_read_tokens`     |
| `usage.cacheWriteInputTokens`                 | `cache_creation_tokens` |
| `usage.input_tokens` (InvokeModel/Anthropic)  | `prompt_tokens`         |
| `usage.output_tokens` (InvokeModel/Anthropic) | `completion_tokens`     |
| `inputTextTokenCount` (Titan)                 | `prompt_tokens`         |
| `results[0].tokenCount` (Titan)               | `completion_tokens`     |
| `meta.billed_units.input_tokens` (Cohere)     | `prompt_tokens`         |
| `meta.billed_units.output_tokens` (Cohere)    | `completion_tokens`     |