Quickstart (5-Minute Guide)
Create your first working LLM program in Haskell with Ollama, OpenAI, or Gemini.
1. Minimal Working Example (Ollama)
Make sure you have Ollama running locally (ollama run llama3.2 or ollama run qwen2.5:7b):
{-# LANGUAGE OverloadedStrings #-}
module Main where
import Control.Monad.Except (runExceptT)
import qualified Data.Text.IO as TIO
import Langchain.Prelude
import Langchain.Provider.Ollama
main :: IO ()
main = do
-- 1. Initialize the ChatModel instance
model <- newOllama "llama3.2" defaultConfig
-- 2. Construct messages
let messages =
[ systemMessage "You are a concise, helpful Haskell tutor."
, userMessage "What is the difference between Functor, Applicative, and Monad in one sentence?"
]
-- 3. Invoke the model
res <- runExceptT $ invoke model messages Nothing
case res of
Left err -> putStrLn ("Invocation Error: " ++ show err)
Right responseMsg -> do
putStrLn "--- LLM Response ---"
TIO.putStrLn (extractMessageText responseMsg)Execution Output
--- LLM Response ---
A Functor maps a normal function over wrapped values, an Applicative maps a wrapped function over wrapped values, and a Monad feeds a wrapped value into a function that returns a wrapped value.
2. Using OpenAI or Gemini
Switching between model providers requires changing only the model constructor:
-- OpenAI
import Langchain.Provider.OpenAI
let model = newOpenAI "sk-..." "gpt-4o"
-- Google Gemini (Interactions v1 API)
import Langchain.Provider.Gemini
let model = newGemini "AIzaSy..." "gemini-3.6-flash" Nothing3. Real-Time Streaming Output
langchain-hs provides conduit-based reactive event streaming with StreamEvent:
import Langchain.Prelude
import Langchain.Provider.Ollama
streamExample :: IO ()
streamExample = do
model <- newOllama "llama3.2" defaultConfig
let messages = [userMessage "Explain Monads briefly."]
-- Stream tokens directly to stdout as they arrive
streamModel model messages $ \event ->
case event of
LLMStart -> putStrLn "[Stream Started]"
LLMChunk chunkText -> putStr (show chunkText)
LLMEnd usage -> putStrLn "\n[Stream Finished]"
_ -> pure ()4. Pure AST Pipelines (RunnableTree)
Instead of evaluating LLM calls eagerly, build a declarative AST and interpret it:
pipeline :: RunnableTree IO Text Text
pipeline =
runLambda (\q -> "Answer in 1 sentence: " <> q)
|>> invokeLLM model
|>> runLambda extractMessageText
runIt :: IO ()
runIt = do
answer <- interpret pipeline "Why is Haskell great for AI?"
putStrLn (show answer)
What’s Next?
Learn how to equip LLMs with tools and autonomous reasoning in the First Agent Guide, or build stateful graph workflows in the First StateGraph Guide. You can also view package releases on Hackage.