How do you configure an LLM in LangChain for text generation?

What you need to know

There are two ways to create a chat model, and they return the same kind of object.

1from langchain.chat_models import init_chat_model2from langchain_openai import ChatOpenAI34# Provider-neutral: the provider comes from the string5llm = init_chat_model(6    "openai:gpt-5.4-mini",7    temperature=0,     # steady output for extraction/classification8    max_tokens=500,    # caps reply length and cost9    timeout=30,        # seconds; never wait forever10    max_retries=2,     # retries 429s, 5xx and network errors with backoff11)1213# Provider-specific: exposes every provider option14llm2 = ChatOpenAI(model="gpt-5.4-mini", temperature=0, max_tokens=500,15                  timeout=30, max_retries=2)1617msg = llm.invoke("Summarise this release note in two sentences: ...")18print(msg.text)                               # the reply as a string19print(msg.usage_metadata)                     # input/output/total tokens20print(msg.response_metadata.get("finish_reason"))

init_chat_model needs the provider package installed (langchain-openai here). Use the provider class when you need a setting only that provider has, such as reasoning_effort on OpenAI reasoning models.

The settings that matter

Reading the reply

invoke returns an AIMessage. In LangChain 1.x, .text gives the reply as one string even when the provider returns a list of content blocks (text plus reasoning or citations); .content is the raw field. finish_reason (OpenAI's name; Anthropic calls it stop_reason) tells you whether the model stopped naturally or hit max_tokens. Use llm.stream(...) to get the reply chunk by chunk for a chat UI.

A real-life example

An invoice extraction service at a Mumbai accounting firm used the provider's default settings. Two problems showed up in the first week: the same invoice gave a different total format on re-runs, and one slow provider response held a worker for 4 minutes. The fix was three arguments. temperature=0 made output stable, max_tokens=400 stopped the occasional rambling reply (the JSON they need is about 150 tokens), and timeout=30 with max_retries=2 meant a stuck call failed in under 2 minutes in the worst case and went to a retry queue. Logging usage_metadata per invoice showed an average cost they could quote to finance.

Follow-up questions to expect

• "Where should the API key go?" — In an environment variable such as OPENAI_API_KEY, which the class reads automatically; pass api_key= only from a secret manager, never a literal.
• "How do you know a reply was cut off?" — Check response_metadata: OpenAI sets finish_reason to length, and Anthropic sets stop_reason to max_tokens, when the limit was hit.
• "Does temperature 0 make output fully deterministic?" — No. It makes it much more stable, but providers do not guarantee identical output, so tests should check shape and meaning, not exact strings.