Langchain Upgrade Migration

LangChain 1.0 Upgrade Migration (Python)

Overview

The first deploy after pip install -U langchain crashes on import with:

ImportError: cannot import name 'ChatOpenAI' from 'langchain.chat_models'

Fix the import, restart, and the next error lands:

ImportError: cannot import name 'LLMChain' from 'langchain.chains'
AttributeError: module 'langchain.agents' has no attribute 'initialize_agent'
AttributeError: 'ConversationBufferMemory' object has no attribute 'save_context'

LangChain 1.0 removed four entire public-API surfaces in one release:

• Provider imports under langchain.chat_models / langchain.llms (pain code P38).
• The LLMChain family under langchain.chains (P39).
• ConversationBufferMemory and siblings under langchain.memory (P40).
• initialize_agent under langchain.agents (P41).

Anything that inspected intermediate_steps also breaks because the tuple shape changed from (AgentAction, observation) to (ToolCall, observation) (P42).

This skill walks a reversible, phased migration:

1. A pre-flight grep audit.
2. A pinned package upgrade (including the langchain-anthropic 1.0 peer-pin against anthropic >= 0.40, P66).
3. Codemod patterns for the seven removed APIs.
4. A rollout playbook with shadow traffic and a sub-five-minute rollback.

It covers 7 named breaking changes and typically touches 10–100 files in a mid-sized service.

The fix for the error above:

# BEFORE (0.3)
from langchain.chat_models import ChatOpenAI

# AFTER (1.0)
from langchain_openai import ChatOpenAI

See codemod-patterns.md for the other six patterns.

Prerequisites

• Python 3.10+ (LangChain 1.0 dropped 3.8/3.9).
• A working test suite for the service being migrated (the playbook runs pytest -W error::DeprecationWarning at every phase).
• Git on a clean working tree — the migration uses per-module commits so rollback is per-commit.
• Access to staging traffic or a request-mirror. Phase 4 of the playbook needs real-shape traffic.
• If conversations are persisted (Redis / Postgres / DynamoDB), a snapshot of the chat-history store before Phase 2. The LangGraph checkpointer uses a new schema and a naive rollback is data-lossy.

Instructions

Step 1 — Pre-flight grep audit

Inventory every 0.3 usage before touching a requirements.txt. Each grep below maps to one pain code and one codemod pattern.

grep -rn "from langchain\.chat_models\|from langchain\.llms" --include="*.py" .          # P38
grep -rn "from langchain\.chains\b\|\bLLMChain\b\|\bRetrievalQA\b" --include="*.py" .    # P39
grep -rn "from langchain\.memory\|ConversationBufferMemory" --include="*.py" .           # P40
grep -rn "initialize_agent\|AgentType\." --include="*.py" .                              # P41
grep -rn "\.tool_input\b\|intermediate_steps" --include="*.py" .                         # P42
grep -rn "astream_log\b" --include="*.py" .                                              # P67

Pipe the full set into langchain-0.3-hits.txt — that file is the migration work list. The migration-detection.md reference has the one-shot bundled block and a line-count triage table.

Step 2 — Pin and upgrade packages together

LangChain 1.0 spans six coordinated packages. A partial upgrade (e.g. pip install -U langchain-anthropic without bumping anthropic) triggers AttributeError at import time (P66). Update all six in the same commit:

langchain>=1.0,<2
langchain-core>=0.3,<0.4
langchain-openai>=1.0
langchain-anthropic>=1.0
langgraph>=1.0,<2
anthropic>=0.40,<1

Apply:

pip install -U \
  "langchain>=1.0,<2" \
  "langchain-core>=0.3,<0.4" \
  "langchain-openai>=1.0" \
  "langchain-anthropic>=1.0" \
  "langgraph>=1.0,<2" \
  "anthropic>=0.40,<1"

Then snapshot the prior state for the rollback: pip freeze > requirements.lock.pre-1.0.txt.

Step 3 — Codemod the four removed APIs

Work through the hits from Step 1 in this order (lowest blast radius first):

1. Provider imports (P38) — mechanical find/replace. from langchain.chat_models import ChatOpenAI → from langchain_openai import ChatOpenAI. Same pattern for ChatAnthropic, OpenAIEmbeddings, Chroma, etc.
2. LLMChain → LCEL (P39) — replace chain = LLMChain(llm=llm, prompt=prompt) with chain = prompt | llm | StrOutputParser(). Caller changes from chain.run(x=1) to chain.invoke({"x": 1}). If the caller treated the result as a dict, unwrap — invoke returns the string directly.
3. initialize_agent → create_react_agent (P41) — swap the import to from langgraph.prebuilt import create_react_agent. Tools written with Tool(name=..., func=...) still work; prefer the @tool decorator from langchain_core.tools. Agent input becomes {"messages": [("user", "...")]}; the final reply is result["messages"][-1].content.
4. ConversationBufferMemory → LangGraph checkpointer (P40) — swap the memory object for MemorySaver() (dev) or SqliteSaver.from_conn_string(...) (prod). Compile the graph/agent with checkpointer=saver, then pass config={"configurable": {"thread_id": "..."}} on every invoke. The thread_id is the conversation primary key.

Full before/after snippets for all four are in codemod-patterns.md.

Step 4 — Update streaming callers (P67)

astream_log still works in 1.0 but is soft-deprecated. The replacement is astream_events(version="v2"):

# BEFORE
async for patch in chain.astream_log({"input": "hi"}):
    for op in patch.ops:
        if op["op"] == "add" and op["path"].endswith("/streamed_output/-"):
            print(op["value"], end="")

# AFTER
async for event in chain.astream_events({"input": "hi"}, version="v2"):
    if event["event"] == "on_chat_model_stream":
        print(event["data"]["chunk"].content, end="")

Event names in v2: on_chain_start, on_chain_end, on_chat_model_start, on_chat_model_stream, on_chat_model_end, on_tool_start, on_tool_end. The payload under data is typed — chunk is an AIMessageChunk, not a raw string.

Step 5 — Fix intermediate_steps consumers (P42)

If any code iterates result["intermediate_steps"] and reads .tool / .tool_input, it breaks silently in 1.0 — the tuples now hold ToolCall dicts, not AgentAction objects. The 1.0 equivalent reads from graph state:

# BEFORE
for action, observation in result["intermediate_steps"]:
    log(action.tool, action.tool_input, observation)

# AFTER
for msg in result["messages"]:
    for tc in getattr(msg, "tool_calls", []) or []:
        log(tc["name"], tc["args"])   # .tool -> "name", .tool_input -> "args"

ToolCall dict keys are name, args, id. There is no tool or tool_input accessor anywhere in 1.0.