feat: Add support for langchain-anthropic in LLMEndpoint

This commit adds support for the langchain-anthropic language model in the LLMEndpoint class. The  function has been updated to include the new model names. Now, when configuring the LLMEndpoint, if the model name starts with claude, the ChatAnthropic class from langchain-anthropic will be used instead of ChatOpenAI from langchain-openai.

backend/core/pyproject.toml

@@ -18,6 +18,7 @@ aiofiles = ">=23.0.0,<25.0.0"
 faiss-cpu = { version = "^1.8.0.post1", optional = true }
 langchain-community = { version = "^0.2.6", optional = true }
 langchain-openai = { version = "^0.1.14", optional = true }
+langchain-anthropic = { version = "^0.1.22", optional = true }
 # To install unstructured, you’ll also need to install the following system dependencies:
 # libmagic, poppler, libreoffice, pandoc, and tesseract.
 # NOTE: for now poetry doesn't support groups as extra:
@@ -49,7 +50,7 @@ unstructured = { version = "^0.15.0", optional = true, extras = [
 docx2txt = { version = "^0.8", optional = true }
 
 [tool.poetry.extras]
-base = ["langchain-community", "faiss-cpu", "langchain-openai"]
+base = ["langchain-community", "faiss-cpu", "langchain-openai", "langchain-anthropic"]
 csv = ["langchain-community"]
 md = ["langchain-community", "unstructured"]
 ipynb = ["langchain-community"]
@@ -66,6 +67,7 @@ all = [
     "unstructured",
     "docx2txt",
     "megaparse",
+    "langchain-anthropic",
 ]
 
 [tool.poetry.group.dev]

backend/core/quivr_core/chat.py

@@ -3,6 +3,7 @@ from typing import Any, Generator, Tuple
 from uuid import UUID, uuid4
 
 from langchain_core.messages import AIMessage, HumanMessage
+
 from quivr_core.models import ChatMessage
 
 

backend/core/quivr_core/llm/llm_endpoint.py

@@ -1,5 +1,6 @@
 from langchain_core.language_models.chat_models import BaseChatModel
 from pydantic.v1 import SecretStr
+
 from quivr_core.brain.info import LLMInfo
 from quivr_core.config import LLMEndpointConfig
 from quivr_core.utils import model_supports_function_calling
@@ -19,13 +20,26 @@ class LLMEndpoint:
     @classmethod
     def from_config(cls, config: LLMEndpointConfig = LLMEndpointConfig()):
         try:
-            from langchain_openai import ChatOpenAI
+            if config.model.startswith("claude"):
+                from langchain_anthropic import ChatAnthropic
 
-            _llm = ChatOpenAI(
-                model=config.model,
-                api_key=SecretStr(config.llm_api_key) if config.llm_api_key else None,
-                base_url=config.llm_base_url,
-            )
+                _llm = ChatAnthropic(
+                    model=config.model,
+                    api_key=SecretStr(config.llm_api_key)
+                    if config.llm_api_key
+                    else None,
+                    base_url=config.llm_base_url,
+                )
+            else:
+                from langchain_openai import ChatOpenAI
+
+                _llm = ChatOpenAI(
+                    model=config.model,
+                    api_key=SecretStr(config.llm_api_key)
+                    if config.llm_api_key
+                    else None,
+                    base_url=config.llm_base_url,
+                )
             return cls(llm=_llm, llm_config=config)
 
         except ImportError as e:

backend/core/quivr_core/utils.py

@@ -1,4 +1,5 @@
 import logging
+import re
 from typing import Any, List, Tuple, no_type_check
 
 from langchain_core.messages import AIMessage, BaseMessage, HumanMessage, SystemMessage
@@ -23,19 +24,19 @@ logger = logging.getLogger("quivr_core")
 
 def model_supports_function_calling(model_name: str):
     models_supporting_function_calls = [
-        "gpt-4",
-        "gpt-4-1106-preview",
-        "gpt-4-0613",
-        "gpt-3.5-turbo-0125",
-        "gpt-3.5-turbo-1106",
-        "gpt-3.5-turbo-0613",
-        "gpt-4-0125-preview",
-        "gpt-3.5-turbo",
-        "gpt-4-turbo",
-        "gpt-4o",
-        "gpt-4o-mini",
+        r"gpt-4(-\d{4}-preview)?",
+        r"gpt-4-\d{4}",
+        r"gpt-3\.5-turbo(-\d{4})?",
+        r"gpt-4-turbo",
+        r"gpt-4o(-mini)?",
+        r"mistral-(small|large)-latest",
+        r"claude-3-opus-\d{8}",
+        r"claude-3-haiku-\d{8}",
+        r"claude-3-sonnet-\d{8}",
     ]
-    return model_name in models_supporting_function_calls
+    return any(
+        re.match(pattern, model_name) for pattern in models_supporting_function_calls
+    )
 
 
 def format_history_to_openai_mesages(

backend/core/tests/test_quivr_rag.py

@@ -1,6 +1,7 @@
 from uuid import uuid4
 
 import pytest
+
 from quivr_core.chat import ChatHistory
 from quivr_core.config import LLMEndpointConfig, RAGConfig
 from quivr_core.llm import LLMEndpoint
@@ -66,3 +67,50 @@ async def test_quivrqarag(
 
     # Assert whole response makes sense
     assert "".join([r.answer for r in stream_responses]) == full_response
+
+
+@pytest.mark.base
+@pytest.mark.asyncio
+async def test_quivrqarag_claude(
+    mem_vector_store, full_response, mock_chain_qa_stream, openai_api_key
+):
+    # Making sure the model
+    llm_config = LLMEndpointConfig(model="claude-3-sonnet-20240229")
+    llm = LLMEndpoint.from_config(llm_config)
+    rag_config = RAGConfig(llm_config=llm_config)
+    chat_history = ChatHistory(uuid4(), uuid4())
+    rag_pipeline = QuivrQARAG(
+        rag_config=rag_config, llm=llm, vector_store=mem_vector_store
+    )
+
+    stream_responses: list[ParsedRAGChunkResponse] = []
+
+    # Making sure that we are calling the func_calling code path
+    assert rag_pipeline.llm_endpoint.supports_func_calling()
+    async for resp in rag_pipeline.answer_astream(
+        "answer in bullet points. tell me something", chat_history, []
+    ):
+        stream_responses.append(resp)
+
+    assert all(
+        not r.last_chunk for r in stream_responses[:-1]
+    ), "Some chunks before last have last_chunk=True"
+    assert stream_responses[-1].last_chunk
+
+    for idx, response in enumerate(stream_responses[1:-1]):
+        assert (
+            len(response.answer) > 0
+        ), f"Sent an empty answer {response} at index {idx+1}"
+
+    # Verify metadata
+    default_metadata = RAGResponseMetadata().model_dump()
+    assert all(
+        r.metadata.model_dump() == default_metadata for r in stream_responses[:-1]
+    )
+    last_response = stream_responses[-1]
+    # TODO(@aminediro) : test responses with sources
+    assert last_response.metadata.sources == []
+    assert last_response.metadata.citations == []
+
+    # Assert whole response makes sense
+    assert "".join([r.answer for r in stream_responses]) == full_response

backend/core/tests/test_utils.py

@@ -13,7 +13,18 @@ from quivr_core.utils import (
 
 def test_model_supports_function_calling():
     assert model_supports_function_calling("gpt-4") is True
+    assert model_supports_function_calling("gpt-4-turbo") is True
+    assert model_supports_function_calling("gpt-4o") is True
+    assert model_supports_function_calling("gpt-4o-mini") is True
     assert model_supports_function_calling("ollama3") is False
+    assert model_supports_function_calling("mistral-small-latest") is True
+    assert model_supports_function_calling("mistral-large-latest") is True
+    assert model_supports_function_calling("gpt-3.5-turbo") is True
+    assert model_supports_function_calling("gpt-3.5-turbo-0125") is True
+
+    assert model_supports_function_calling("gpt-8.4") is False
+    assert model_supports_function_calling("mistral-burger-latest") is False
+    assert model_supports_function_calling("claude-3-opus-20240229") is True
 
 
 def test_get_prev_message_incorrect_message():