diff --git a/CHANGELOG.md b/CHANGELOG.md --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -1,5 +1,13 @@ # Changelog for langchain-hs-gemini +## 0.0.6.2 - 2026-11-14 + +> **Note**: `0.0.6.1` was skipped globally after `langchain-hs-ollama-v0.0.6.1`. + +- Added incremental tool-call emission during SSE streaming in `Langchain.Provider.Gemini` (`ChatModel.stream`). +- Consolidates Interactions v1 Servant bindings (`Gemini.Client.*`) and `Langchain.Embeddings.Gemini` under aligned `0.0.6.2`. +- Updated `langchain-hs-core` dependency bound to `>= 0.0.6.2 && < 0.0.7`. + ## 0.0.6.0 - 2026-10-02 - Initial release. Split from the monolithic `langchain-hs` package. diff --git a/langchain-hs-gemini.cabal b/langchain-hs-gemini.cabal --- a/langchain-hs-gemini.cabal +++ b/langchain-hs-gemini.cabal @@ -5,7 +5,7 @@ -- see: https://github.com/sol/hpack name: langchain-hs-gemini -version: 0.0.6.0 +version: 0.0.6.2 synopsis: Google Gemini provider for langchain-hs description: Google Gemini ChatModel provider for langchain-hs. Supports multi-modal content, streaming via Server-Sent Events, and function calling. Does not depend on OpenAI or Ollama packages. category: AI @@ -51,7 +51,7 @@ , http-client ==0.7.* , http-client-tls >=0.3 && <0.5 , http-conduit ==2.* - , langchain-hs-core ==0.0.6.* + , langchain-hs-core >=0.0.6.2 && <0.0.7 , mtl >=2.2 && <2.4 , resourcet >=1.2 && <1.4 , servant >=0.19 && <0.21 diff --git a/src/Langchain/Provider/Gemini.hs b/src/Langchain/Provider/Gemini.hs --- a/src/Langchain/Provider/Gemini.hs +++ b/src/Langchain/Provider/Gemini.hs @@ -29,14 +29,16 @@ ) where import Control.Applicative ((<|>)) +import Control.Monad (forM_) import Control.Monad.Except (throwError) import Control.Monad.IO.Class (liftIO) -import Data.Aeson (FromJSON, ToJSON, decodeStrict, object, toJSON, (.=)) +import Data.Aeson (FromJSON, ToJSON, decodeStrict, eitherDecodeStrict, object, toJSON, (.=)) import Data.Conduit (await, yield, (.|)) import qualified Data.List as List import qualified Data.List.NonEmpty as NonEmpty import qualified Data.Map.Strict as Map import Data.Maybe (fromMaybe) +import qualified Data.Set as Set import Data.Text (Text) import qualified Data.Text as T import qualified Data.Text.Encoding as TE @@ -270,31 +272,101 @@ Nothing -> newGeminiClientEnv (apiKey provider) let req = buildCreateRequest provider msgs mbConfig - (accumulated, mbCompletedInter) <- - streamGeminiInteraction env req .| consumeStream rId "" Nothing + (accumulated, pendingCalls, emittedCalls, thoughtSigs, mbCompletedInter) <- + streamGeminiInteraction env req .| consumeStream rId "" Map.empty [] [] Nothing + -- Finalize any tool calls whose step.start was not followed by step.stop + remainingCalls <- traverse finalizePendingCall (Map.elems pendingCalls) + forM_ remainingCalls $ \tc -> + yield $ LLMChunk rId "" (Just tc) + + let allEmitted = reverse emittedCalls ++ remainingCalls case mbCompletedInter of Just completedInter -> do let finalMsg = interactionToMessage completedInter usage = toTokenUsage (interactionUsage completedInter) + alreadyEmittedIds = Set.fromList [(toolCallId tc, toolCallName tc) | tc <- allEmitted] + missingChunkCalls = + filter + (\tc -> not $ Set.member (toolCallId tc, toolCallName tc) alreadyEmittedIds) + (fromMaybe [] (messageToolCalls finalMsg)) + forM_ missingChunkCalls $ \tc -> + yield $ LLMChunk rId "" (Just tc) yield $ LLMEnd rId finalMsg usage Nothing -> do - let finalMsg = assistantMessage accumulated + let base = assistantMessage accumulated + withCalls = + if null allEmitted + then base + else base {messageToolCalls = Just allEmitted} + sigs = reverse thoughtSigs + finalMsg = + if null sigs + then withCalls + else + withCalls + { messageMetadata = + Map.insert + "langchain.gemini.thoughtSignatures" + (toJSON sigs) + (messageMetadata withCalls) + } yield $ LLMEnd rId finalMsg Nothing where - consumeStream rId acc mbCompleted = + finalizePendingCall (cid, cname, initArgs, argText) + | T.null (T.strip argText) = + pure + ToolCall + { toolCallId = cid + , toolCallType = "function" + , toolCallName = cname + , toolCallArguments = initArgs + } + | otherwise = + case eitherDecodeStrict (TE.encodeUtf8 argText) of + Left err -> + throwError $ llmError ("Invalid JSON arguments in Gemini tool call: " <> T.pack err) Nothing Nothing + Right parsedArgs -> + pure + ToolCall + { toolCallId = cid + , toolCallType = "function" + , toolCallName = cname + , toolCallArguments = parsedArgs + } + + consumeStream rId acc pending emitted sigs mbCompleted = await >>= \case - Nothing -> pure (acc, mbCompleted) + Nothing -> pure (acc, pending, emitted, sigs, mbCompleted) + Just (SseStepStart idx FunctionCallStep {..}) -> + let pending' = Map.insert idx (callId, callName, callArguments, "") pending + in consumeStream rId acc pending' emitted sigs mbCompleted Just (SseStepDelta _ (TextDelta txt)) -> do yield $ LLMChunk rId txt Nothing - consumeStream rId (acc <> txt) mbCompleted + consumeStream rId (acc <> txt) pending emitted sigs mbCompleted + Just (SseStepDelta idx (ArgumentsDelta argDelta)) -> + let pending' = + Map.adjust + (\(cid, cname, initArgs, argTxt) -> (cid, cname, initArgs, argTxt <> argDelta)) + idx + pending + in consumeStream rId acc pending' emitted sigs mbCompleted + Just (SseStepDelta _ (ThoughtSignatureDelta sig)) -> + consumeStream rId acc pending emitted (sig : sigs) mbCompleted + Just (SseStepStop idx) -> + case Map.lookup idx pending of + Nothing -> consumeStream rId acc pending emitted sigs mbCompleted + Just entry -> do + tc <- finalizePendingCall entry + yield $ LLMChunk rId "" (Just tc) + consumeStream rId acc (Map.delete idx pending) (tc : emitted) sigs mbCompleted Just (SseInteractionCompleted inter) -> - consumeStream rId acc (Just inter) + consumeStream rId acc pending emitted sigs (Just inter) Just (SseError code msg) -> throwError $ llmError (code <> ": " <> msg) Nothing Nothing Just _ -> - consumeStream rId acc mbCompleted + consumeStream rId acc pending emitted sigs mbCompleted toTokenUsage Nothing = Nothing toTokenUsage (Just Usage {..}) = diff --git a/test/Test/Langchain/Provider/GeminiSpec.hs b/test/Test/Langchain/Provider/GeminiSpec.hs --- a/test/Test/Langchain/Provider/GeminiSpec.hs +++ b/test/Test/Langchain/Provider/GeminiSpec.hs @@ -163,6 +163,58 @@ case (events, reverse events) of (LLMStart {} : _, LLMEnd {} : _) -> pure () other -> assertFailure ("Unexpected events: " ++ show other) + , testCase "stream emits toolCallDelta chunks from step.start/step.delta/step.stop" $ do + let sseBody = + "data: {\"event_type\": \"step.delta\", \"index\": 0, \"delta\": {\"type\": \"text\", \"text\": \"Checking \"}}\n\n" + <> "data: {\"event_type\": \"step.start\", \"index\": 1, \"step\": {\"type\": \"function_call\", \"id\": \"call_g1\", \"name\": \"get_weather\", \"arguments\": {}}}\n\n" + <> "data: {\"event_type\": \"step.delta\", \"index\": 1, \"delta\": {\"type\": \"arguments\", \"arguments\": \"{\\\"city\\\":\\\"Paris\\\"}\"}}\n\n" + <> "data: {\"event_type\": \"step.stop\", \"index\": 1}\n\n" + <> "data: {\"event_type\": \"interaction.completed\", \"interaction\": {\"id\": \"v1_tc\", \"status\": \"requires_action\", \"steps\": [{\"type\": \"model_output\", \"content\": [{\"type\": \"text\", \"text\": \"Checking \"}]}, {\"type\": \"function_call\", \"id\": \"call_g1\", \"name\": \"get_weather\", \"arguments\": {\"city\": \"Paris\"}}]}}\n\n" + app _req respond = + respond $ + responseStream status200 [("Content-Type", "text/event-stream")] $ \write flush -> do + write (Builder.lazyByteString sseBody) + flush + expectedTc = + ToolCall + { toolCallId = "call_g1" + , toolCallType = "function" + , toolCallName = "get_weather" + , toolCallArguments = [aesonQQ|{"city": "Paris"}|] + } + withMockServer app $ \url -> do + let provider = newGemini "fake-key" "gemini-3.6-flash" (Just url) + eEvents <- + runResourceT $ + runExceptT $ + collectEvents $ + stream provider [userMessage "Weather in Paris?"] Nothing + case eEvents of + Left err -> assertFailure (show err) + Right events -> do + let toolChunks = [tc | LLMChunk _ _ (Just tc) <- events] + toolChunks @?= [expectedTc] + case reverse events of + LLMEnd _ finalMsg _ : _ -> + messageToolCalls finalMsg @?= Just [expectedTc] + _ -> assertFailure ("Expected LLMEnd, got: " ++ show events) + , testCase "stream emits toolCallDelta from interaction.completed fallback" $ do + let sseBody = + "data: {\"event_type\": \"interaction.completed\", \"interaction\": {\"id\": \"v1_tc2\", \"status\": \"requires_action\", \"steps\": [{\"type\": \"function_call\", \"id\": \"call_g2\", \"name\": \"ping\", \"arguments\": {}}]}}\n\n" + app _req respond = + respond $ + responseStream status200 [("Content-Type", "text/event-stream")] $ \write flush -> do + write (Builder.lazyByteString sseBody) + flush + expectedTc = ToolCall "call_g2" "function" "ping" [aesonQQ|{}|] + withMockServer app $ \url -> do + let provider = newGemini "fake-key" "gemini-3.6-flash" (Just url) + eEvents <- runResourceT $ runExceptT $ collectEvents $ stream provider [userMessage "Ping"] Nothing + case eEvents of + Left err -> assertFailure (show err) + Right events -> do + let toolChunks = [tc | LLMChunk _ _ (Just tc) <- events] + toolChunks @?= [expectedTc] ] , testGroup "Embeddings GeminiEmbeddings (Mock)"