• This is actually a bug report.
• I am not getting good LLM Results
• [] I have tried asking for help in the community on discord or discussions and have not received a response.
• [] I have tried searching the documentation and have not found an answer.

What Model are you using?

• gpt-3.5-turbo
• gpt-4-turbo
• gpt-4
• Other (please specify): claude-sonnet-4-6 via AWS Bedrock (anthropic.AnthropicBedrock)

Describe the bug

When a Pydantic validation error occurs during a structured extraction call using ANTHROPIC_TOOLS or ANTHROPIC_REASONING_TOOLS mode via AWS Bedrock, the retry mechanism (reask_anthropic_tools) fails on every subsequent attempt with HTTP 400. The error is:

Error code: 400 - {'message': 'messages.1.content.1.tool_use.caller: Input should be a valid dictionary or object to extract fields from'}

Root cause: anthropic==0.88.0 added a new caller: Optional[Caller] = None field to ToolUseBlock. Bedrock does not populate this field — it returns caller=None. In reask_anthropic_tools, the failed completion is serialized via plain model_dump(), which produces "caller": null in the assistant message. The Anthropic API rejects null for this field.

# providers/anthropic/utils.py — reask_anthropic_tools, line 168
for content in response.content:
    assistant_content.append(content.model_dump())  # ← includes "caller": null for Bedrock responses

Why it only affects Bedrock: The direct Anthropic API correctly populates caller=DirectCaller(type='direct') on tool use responses, so model_dump() produces a valid dict. Bedrock returns caller=None, so model_dump() produces "caller": null, which the API rejects.

Verified with live calls to both providers using identical code and the same impossible validation constraint (value: int = Field(ge=10, le=5)) to force retries:

# Direct Anthropic API — generation 2 response:
ToolUseBlock(caller=DirectCaller(type='direct'), ...)
# model_dump() → {"caller": {"type": "direct"}, ...}  ✓ retries succeed (Pydantic errors only)

# Bedrock — generation 2 response:
ToolUseBlock(caller=None, ...)
# model_dump() → {"caller": null, ...}  ✗ generation 3 fails with HTTP 400

On direct Anthropic API, all 3 generations retry successfully and only fail due to the impossible Pydantic constraint — proving the retry mechanism itself works when caller is populated. On Bedrock, generation 3 immediately returns 400.

Both ANTHROPIC_TOOLS and ANTHROPIC_REASONING_TOOLS are affected since they share the same reask_anthropic_tools handler (see ANTHROPIC_HANDLERS registry in the same file).

To Reproduce

import instructor
import anthropic
from pydantic import BaseModel, Field

class MyExtraction(BaseModel):
    summary: str
    value: int = Field(ge=10, le=5)  # impossible constraint — forces validation failure on every generation

bedrock_client = instructor.from_anthropic(
    anthropic.AnthropicBedrock(
        aws_access_key="...",
        aws_secret_key="...",
        aws_region="us-east-1",
    ),
    mode=instructor.Mode.ANTHROPIC_REASONING_TOOLS,
)

try:
    response = bedrock_client.create(
        model="us.anthropic.claude-sonnet-4-6",
        max_tokens=4000,
        thinking={"type": "enabled", "budget_tokens": 2000},
        messages=[{"role": "user", "content": "Extract: some text"}],
        response_model=MyExtraction,
    )
except Exception as e:
    print(type(e).__name__, str(e))

What happens step by step:

1. Generation 1: Model notices the contradictory schema and refuses to call the tool → list should have at least 1 item validation error (no ToolUseBlock yet).
2. Generation 2: After reask, model calls the tool with value=10 → Pydantic ge=10, le=5 validation fails. Bedrock returns ToolUseBlock(caller=None, ...).
3. reask_anthropic_tools serializes the generation 2 response via content.model_dump(), producing {"caller": null, ...} in the assistant message.
4. Generation 3: API immediately returns HTTP 400 — all further retries fail identically.

Actual reproduction output (confirmed on AWS Bedrock, eu-west-1):

Generation 2 completion showing caller=None:

ToolUseBlock(
    id='toolu_bdrk_015F3ce6KTnAWmxzcMFoGtdz',
    caller=None,   # ← Bedrock does not populate this field
    input={'summary': 'some text', 'value': 10},
    name='MyExtraction',
    type='tool_use'
)

Generation 3 exception:

Error code: 400 - {'message': 'messages.3.content.1.tool_use.caller: Input should be a valid dictionary or object to extract fields from'}

Expected behavior

After a Pydantic validation error, the retry should succeed. The assistant message sent back to the API should not include "caller": null. The caller field should be omitted when it is None so the API accepts the retry message.

Screenshots

N/A — error is reproduced programmatically.

Proposed fix

In providers/anthropic/utils.py, reask_anthropic_tools, line 168 — use exclude_none=True:

# Before:
assistant_content.append(content.model_dump())

# After:
assistant_content.append(content.model_dump(exclude_none=True))

This strips "caller": null from Bedrock responses while leaving direct Anthropic API responses unchanged (where caller is a valid DirectCaller dict and would be preserved).

Environment:

• instructor==1.15.1
• anthropic==0.88.0
• Python 3.13
• AWS Bedrock (AnthropicBedrock)