Feature Tutorials

Testing & Mocking Infrastructure

Write deterministic pure unit tests and CI pipelines without requiring a running Ollama server.

Why Mock Testing for LLMs?

Running integration tests against a live LLM daemon during CI is often slow, non-deterministic, and requires heavy GPU compute.

ollama-haskell includes a built-in mock testing module (Ollama.Testing) that lets you test your business logic against mocked LLM responses with zero network overhead.

1. Creating a Mock Client

Use newMockClient or withMockClient:

{-# LANGUAGE OverloadedStrings #-}
module Main where

import Data.Aeson (encode)
import Data.ByteString.Lazy qualified as BSL
import Data.List.NonEmpty (NonEmpty ((:|)))
import Ollama
import Ollama.Testing
import Test.Tasty
import Test.Tasty.HUnit

-- Sample business logic that interacts with Ollama
askAssistant :: OllamaClient -> IO (Either OllamaError String)
askAssistant client = do
  let req = chatRequest "qwen3.5:2b" (userMessage "Ping" :| [])
  res <- chat client req
  pure $ fmap (maybe "" (show . messageContent) . crMessage) res

testMockChat :: TestTree
testMockChat = testCase "askAssistant returns mocked response" $ do
  -- 1. Create mock client with predefined serialized response
  let mockResp = mockChatResponse "qwen3.5:2b" "Pong from mock!"
  client <- newMockClient (BSL.toStrict (encode mockResp))

  -- 2. Execute business logic
  result <- askAssistant client

  -- 3. Assert deterministic outcome
  case result of
    Right content -> content @?= "\"Pong from mock!\""
    Left err      -> assertFailure $ "Unexpected error: " <> show err

main :: IO ()
main = defaultMain testMockChat

2. Available Mock Helpers

Helper Function	Target Endpoint	Description
`mockChatResponse`	`chat`	Mocks chat completion responses with custom `Message`
`mockGenerateResponse`	`generate`	Mocks text generation with custom string response
`mockEmbedResponse`	`embed`	Mocks embedding vectors
`mockListModelsResponse`	`listModels`	Mocks list of installed local models

All mock clients are fully thread-safe and can be used directly in test suites across multiple threads.