langchain-model-inference

Imported from jeremylongshore/tons-of-skills-marketplace (plugins/saas-packs/langchain-py-pack/skills/langchain-model-inference/SKILL.md). Install upstream with npx skills add jeremylongshore/tons-of-skills-marketplace --skill langchain-model-inference. Copyright stays with the author (MIT).

LangChain Model Inference (Python)

Overview

AIMessage.content is a str on simple OpenAI calls and a list[dict] on Claude the instant any tool_use, thinking, or image block enters the response. Code that does message.content.lower() crashes with AttributeError: 'list' object has no attribute 'lower' — the #1 first-production-call LangChain 1.0 bug on Anthropic. And that is one of four separate "content shape" pitfalls in this skill:

• P02 — AIMessage.content list-vs-string divergence
• P03 — with_structured_output(method="function_calling") silently drops Optional[list[X]] fields on ~40% of real schemas
• P05 — temperature=0 is not deterministic on Anthropic even though it is on OpenAI
• P58 — Claude expects the system message at position 0; middleware that reorders messages makes it silently ignored

This skill walks through ChatAnthropic, ChatOpenAI, and ChatGoogleGenerativeAI initialization; model routing; token counting that is actually correct during streaming; content-block iteration; and a decision tree for with_structured_output methods that holds up on real schemas. Pin: langchain-core 1.0.x, langchain-anthropic 1.0.x, langchain-openai 1.0.x, langchain-google-genai 1.0.x. Pain-catalog anchors: P01, P02, P03, P04, P05, P53, P54, P58, P63, P64, P65.

Prerequisites

• Python 3.10+
• langchain-core >= 1.0, < 2.0
• At least one provider package: pip install langchain-anthropic langchain-openai
• Provider API key(s): ANTHROPIC_API_KEY, OPENAI_API_KEY, GOOGLE_API_KEY

Instructions

Step 1 — Initialize a chat model with explicit, version-safe defaults

from langchain_anthropic import ChatAnthropic
from langchain_openai import ChatOpenAI

claude = ChatAnthropic(
    model="claude-sonnet-4-6",
    temperature=0,
    max_tokens=4096,
    timeout=30,       # seconds. Default is None — hangs forever on provider stall.
    max_retries=2,    # Retries, not attempts. See P30 in pain catalog.
)

gpt4o = ChatOpenAI(
    model="gpt-4o",
    temperature=0,
    timeout=30,
    max_retries=2,
)

Explicit timeout and max_retries are not optional in production — the defaults are wrong for every workload we have measured. max_retries=6 (the default on ChatOpenAI) means a single logical call can bill as 7 requests on flaky networks.

Step 2 — Iterate AIMessage.content as typed blocks, not strings

from langchain_core.messages import AIMessage

def extract_text(msg: AIMessage) -> str:
    """Safe on both provider shapes. Works for streaming deltas too.

    Handles both dict blocks (provider-native) and typed block objects
    (LangChain 1.0 wrappers) — which Gemini, OpenAI tools, and future
    SDK versions may return.
    """
    if isinstance(msg.content, str):
        return msg.content
    parts = []
    for block in msg.content:
        # Block may be a dict (provider-native) or a typed object (1.0 wrapper)
        block_type = block.get("type") if isinstance(block, dict) else getattr(block, "type", None)
        if block_type == "text":
            parts.append(block["text"] if isinstance(block, dict) else block.text)
    return "".join(parts)

AIMessage.text() (1.0+) does this for you in most cases — prefer it. Roll your own only when you need to filter by block type (tool_use, image, thinking). See Content Blocks for the full block-type reference and streaming-delta shape.

Step 3 — Route across providers with a factory, not a conditional

from langchain_core.language_models import BaseChatModel

# Version-safe defaults applied to every model the factory builds.
# Callers can override via **kwargs.
_SAFE_DEFAULTS = {"timeout": 30, "max_retries": 2}

def chat_model(provider: str, **kwargs) -> BaseChatModel:
    defaults = {**_SAFE_DEFAULTS, **kwargs}  # caller's kwargs win
    if provider == "anthropic":
        return ChatAnthropic(model="claude-sonnet-4-6", **defaults)
    if provider == "openai":
        return ChatOpenAI(model="gpt-4o", **defaults)
    if provider == "gemini":
        from langchain_google_genai import ChatGoogleGenerativeAI
        return ChatGoogleGenerativeAI(model="gemini-2.5-pro", **defaults)
    raise ValueError(f"Unknown provider: {provider!r}")

A factory centralizes the version-safe defaults from Step 1 (timeout=30, max_retries=2) and the structured-output method pick from Step 5. Chains depend on the BaseChatModel protocol, not the concrete class. Callers override with chat_model("openai", timeout=60) when they need it.

Step 4 — Count tokens correctly during streaming

ChatAnthropic.stream() does not populate response_metadata["token_usage"] until the stream closes (P01). If your cost dashboard reads on_llm_end, it lags by the stream duration. Use astream_events(version="v2"):

async for event in claude.astream_events({"input": "..."}, version="v2"):
    if event["event"] == "on_chat_model_stream":
        chunk = event["data"]["chunk"]
        if hasattr(chunk, "usage_metadata") and chunk.usage_metadata:
            meter.record(chunk.usage_metadata["input_tokens"],
                         chunk.usage_metadata["output_tokens"])

See Token Accounting for per-provider differences (Anthropic reports input/output/cache separately; OpenAI aggregates; Gemini reports completion-only on stream start).

Step 5 — Pick the right with_structured_output method

from pydantic import BaseModel, ConfigDict

class Plan(BaseModel):
    model_config = ConfigDict(extra="ignore")  # P53 — models add helpful extra fields
    steps: list[str]
    estimated_minutes: int

structured = claude.with_structured_output(Plan, method="json_schema")
plan = structured.invoke("Plan a 3-step deploy")

Avoid Optional[list[X]] fields — they silently return None on some providers (P03). See Structured Output Methods for a concrete comparison matrix and fallback pattern.

Output

• Chat models initialized with explicit timeouts (30s) and max_retries=2
• Content-safe extractor that handles both str and list[dict] shapes
• Factory-based routing with a single BaseChatModel return type
• Streaming token counter that reports incrementally, not at stream end
• with_structured_output chosen per provider capability, with Pydantic validation

Error Handling

Examples

Routing: cheap draft, expensive final

A common pattern — draft with gpt-4o-mini, finalize with claude-sonnet-4-6. The factory in Step 3 makes this trivial; combined with with_structured_output the finalize step returns a typed object.

See Provider Quirks for the full draft-then-finalize example including the token budget calculation.

Extracting tool calls from a single-shot response

A classification task that should return one tool call with a typed argument. Use bind_tools([...], tool_choice={"type": "tool", "name": "Classify"}) for a single forced call — but never loop on a forced choice (P63).

See Structured Output Methods for the worked example and the decision tree for tool vs structured-output for extraction.

Multi-modal: screenshot plus prompt

Images are passed as content blocks, but the block shape differs between providers (P64). LangChain 1.0 abstracts this into a universal image content block.

See Content Blocks for the universal shape and per-provider adapter examples.

Resources

• LangChain Python: Chat models
• AIMessage API reference
• with_structured_output
• astream_events v2
• LangChain 1.0 release notes
• Pack pain catalog: docs/pain-catalog.md (entries P01-P05, P53, P54, P58, P63-P65)