LangChain .bind_tools() — provider schema translation failures

from typing import List, Literal
from langchain_core.tools import tool
from langchain_openai import ChatOpenAI
from langchain_anthropic import ChatAnthropic
from langchain_google_genai import ChatGoogleGenerativeAI

@tool
def search_docs(
    query: str,
    top_k: int = 5,
    doc_types: List[str] = [],                    # empty list vs Optional
    sort: Literal["relevance", "date"] = "relevance",  # Literal enum
) -> str:
    """Search the internal document index.

    Args:
        query: Search query in plain English.
        top_k: Number of results to return (1-20).
        doc_types: Restrict to these types, or [] for all.
        sort: How to order results.
    """
    return f"[{sort}] {top_k} results for {query!r}"

# Same tool binds cleanly to every provider
openai_llm = ChatOpenAI(model="gpt-4o-mini").bind_tools([search_docs])
anthropic_llm = ChatAnthropic(model="claude-sonnet-5").bind_tools([search_docs])
google_llm = ChatGoogleGenerativeAI(model="gemini-2.5-pro").bind_tools([search_docs])

# All three respond with tool_use blocks LangChain understands uniformly
for name, llm in [("openai", openai_llm), ("anthropic", anthropic_llm), ("google", google_llm)]:
    response = llm.invoke("Find recent policy documents")
    tool_calls = response.tool_calls if hasattr(response, "tool_calls") else []
    print(f"{name}: {len(tool_calls)} tool calls")

"""Validate tools bind cleanly to every supported provider."""
import pytest
from langchain_openai import ChatOpenAI
from langchain_anthropic import ChatAnthropic
from langchain_google_genai import ChatGoogleGenerativeAI

from my_app.tools import ALL_TOOLS  # your production tool set

PROVIDERS = [
    ("openai_strict", lambda: ChatOpenAI(model="gpt-4o-mini")),
    ("openai_strict_true", lambda: ChatOpenAI(model="gpt-4o-mini")),
    ("anthropic", lambda: ChatAnthropic(model="claude-sonnet-5")),
    ("google", lambda: ChatGoogleGenerativeAI(model="gemini-2.5-pro")),
]

@pytest.mark.parametrize("provider_name,model_factory", PROVIDERS)
@pytest.mark.parametrize("tool", ALL_TOOLS, ids=lambda t: t.name)
def test_tool_binds_to_provider(provider_name, model_factory, tool):
    """Every tool must bind to every supported provider without error."""
    llm = model_factory()
    kwargs = {}
    if provider_name == "openai_strict_true":
        kwargs["strict"] = True
    try:
        bound = llm.bind_tools([tool], **kwargs)
        # For OpenAI, we can also inspect the serialized schema
        # For all providers, this at least validates the LangChain-side schema
    except Exception as e:
        pytest.fail(f"{tool.name} failed to bind to {provider_name}: {e}")

# Optional: check the schema round-trip against provider validators
def test_openai_schema_shape(tool=search_docs):
    from langchain_core.utils.function_calling import convert_to_openai_tool
    openai_schema = convert_to_openai_tool(tool)
    assert "function" in openai_schema
    assert openai_schema["function"]["description"], "Missing description"

from langchain_openai import ChatOpenAI
from langchain_anthropic import ChatAnthropic

# LangChain-normalized tool_choice — same syntax across providers

# 1) Model must call SOME tool (any of the bound tools)
openai_forced = ChatOpenAI(model="gpt-4o-mini").bind_tools(tools, tool_choice="any")
anthropic_forced = ChatAnthropic(model="claude-sonnet-5").bind_tools(tools, tool_choice="any")

# 2) Model must call a SPECIFIC tool
openai_specific = ChatOpenAI(model="gpt-4o-mini").bind_tools(tools, tool_choice="search_docs")
anthropic_specific = ChatAnthropic(model="claude-sonnet-5").bind_tools(tools, tool_choice="search_docs")

# 3) Model must NOT call any tool (get a text response instead)
openai_no_tools = ChatOpenAI(model="gpt-4o-mini").bind_tools(tools, tool_choice="none")

# 4) Auto — model decides (default; usually best)
openai_auto = ChatOpenAI(model="gpt-4o-mini").bind_tools(tools, tool_choice="auto")

# ❌ AVOID — provider-specific dict forms; not portable
# ChatOpenAI(...).bind_tools(tools, tool_choice={"type": "function", "function": {"name": "search_docs"}})
# ChatAnthropic(...).bind_tools(tools, tool_choice={"type": "tool", "name": "search_docs"})