Merge pull request #171 from mit-submit/development/anthropic_claude

Development/anthropic claude

a2rchi/chains/base.py

@@ -1,5 +1,6 @@
 """Chain for chatting with a vector database."""
 from __future__ import annotations
+from pydantic import BaseModel
 from loguru import logger
 from langchain.callbacks import FileCallbackHandler
 
@@ -12,7 +13,10 @@
 from langchain.chains.llm import LLMChain
 from langchain.schema import BaseRetriever, Document
 from langchain.schema.prompt_template import BasePromptTemplate
+from langchain_core.runnables import RunnableSequence, RunnablePassthrough
 from typing import Any, Dict, List, Optional, Tuple
+from typing import Callable
+# from pydantic import model_rebuild
 import os 
 
 
@@ -46,7 +50,20 @@ class BaseSubMITChain(BaseConversationalRetrievalChain):
     """
     retriever: BaseRetriever # Index to connect to
     max_tokens_limit: Optional[int] = None # restrict doc length to return from store, enforced only for StuffDocumentChain
-    get_chat_history: Optional[function] = _get_chat_history
+    get_chat_history: Optional[Callable[[List[Tuple[str, str]]], str]] = _get_chat_history
+
+    # kuangfei: the rebuild logic is default or rewrite
+    # @classmethod
+    # def model_rebuild(
+    #     cls,
+    #     *,
+    #     force: bool = False,
+    #     raise_errors: bool = True,
+    #     _parent_namespace_depth: int = 2,
+    #     _types_namespace: dict[str, Any] | None = None,
+    # ) -> bool | None:
+    #     return False
+
 
     def _reduce_tokens_below_limit(self, docs: List[Document]) -> List[Document]:
         num_docs = len(docs)
@@ -104,6 +121,17 @@ def from_llm(
             callbacks = [handler],
             verbose=verbose,
         )
+
+        # llm_chain = RunnableSequence(
+        #     {
+        #         "sentence": RunnablePassthrough(),
+        #         "language": RunnablePassthrough()
+        #     }
+        #     | _prompt
+        #     | llm
+        #     # | output_parser
+        # )
+
         doc_chain = StuffDocumentsChain(
             llm_chain=llm_chain,
             document_variable_name=document_variable_name,
@@ -114,7 +142,7 @@ def from_llm(
         condense_question_chain = LLMChain(
             llm=_llm, prompt=condense_question_prompt, callbacks = [handler], verbose=verbose
         )
-
+        
         return cls(
             retriever=retriever,
             combine_docs_chain=doc_chain,

a2rchi/chains/chain.py

@@ -1,7 +1,7 @@
 from a2rchi.chains.base import BaseSubMITChain as BaseChain
 
 from chromadb.config import Settings
-from langchain.vectorstores import Chroma
+from langchain_chroma import Chroma
 
 import chromadb
 import time

a2rchi/chains/models.py

@@ -7,9 +7,15 @@
 from langchain.callbacks.manager import CallbackManagerForLLMRun
 from langchain.llms.base import LLM
 from langchain.chat_models import ChatOpenAI
-from langchain.chat_models import ChatAnthropic
+from langchain_anthropic import ChatAnthropic
 from langchain.llms import LlamaCpp
 
+
+import requests
+from typing import Optional, List
+
+
+
 class BaseCustomLLM(LLM):
     """
     Abstract class used to load a custom LLM
@@ -58,7 +64,6 @@ class AnthropicLLM(ChatAnthropic):
     """
 
     model: str = "claude-3-opus-20240229"
-
     temp: int = 1
 
 
@@ -234,3 +239,77 @@ def __call__(self, output_text):
             report += "|" + "|".join(f"{n:^10}" for n in scores.keys()) + "|\n"
             report += "|" + "|".join(f"{n:^10}" for n in scores.values()) + "|\n"
         return "Salesforce Content Safety Flan T5 Base", is_safe, report
+
+class BaseCustomLLM(LLM):
+    """
+    Abstract class used to load a custom LLM
+    """
+    n_tokens: int = 100 # this has to be here for parent LLM class
+
+    @property
+    def _llm_type(self) -> str:
+        return "custom"
+
+    @abstractmethod
+    def _call(
+        self,
+        prompt: str,
+        stop: Optional[List[str]] = None,
+        run_manager: Optional[CallbackManagerForLLMRun] = None,
+    ) -> str:
+        pass
+
+
+class ClaudeLLM(BaseCustomLLM):
+    """
+    An LLM class that uses Anthropic's Claude model.
+    """
+    #TODO: obscure api key in final production version
+    api_key: str  = "INSERT KEY HERE!!!" # Claude API key
+    base_url: str = "https://api.anthropic.com/v1/messages"  # Anthropic API endpoint
+    model_name: str = "claude-3-5-sonnet-20240620"  # Specify the model version to use
+
+    verbose: bool = False
+
+    def _call(
+        self,
+        prompt: str = None,
+        stop: Optional[List[str]] = None,
+        run_manager: Optional[CallbackManagerForLLMRun] = None,
+        max_tokens: int = 1024,
+    ) -> str:
+
+        if stop is not None:
+            print("WARNING : currently this model does not support stop tokens")
+
+        if self.verbose:
+            print(f"INFO : Starting call to Claude with prompt: {prompt}")
+
+        headers = {
+            "x-api-key": self.api_key,  # Use the API key for the x-api-key header
+            "anthropic-version": "2023-06-01",  # Add the required version header
+            "Content-Type": "application/json"
+        }
+
+        # Modify the payload to match the required structure
+        payload = {
+            "model": self.model_name,  # You can keep this dynamic based on your code
+            "max_tokens": max_tokens,  # Update to match the required max_tokens
+            "messages": [  # Use a list of messages where each message has a role and content
+                {"role": "user", "content": prompt}  # Prompt becomes part of the message content
+            ]
+        }
+
+        if self.verbose:
+            print("INFO: Sending request to Claude API")
+
+        # Send request to Claude API
+        response = requests.post(self.base_url, headers=headers, json=payload)
+
+        if response.status_code == 200:
+            completion = response.json()["content"][0]["text"]
+            if self.verbose:
+                print(f"INFO : received response from Claude API: {completion}")
+            return completion
+        else:
+            raise Exception(f"API request to Claude failed with status {response.status_code}, {response.text}")

a2rchi/utils/config_loader.py

@@ -23,7 +23,7 @@ def load_config(self):
 
             # change the model class parameter from a string to an actual class
             MODEL_MAPPING = {
-                "AnthropicLLM": AnthropicLLM
+                "AnthropicLLM": AnthropicLLM,
                 "OpenAILLM": OpenAILLM,
                 "DumbLLM": DumbLLM,
                 "LlamaLLM": LlamaLLM

a2rchi/utils/scraper.py

@@ -1,4 +1,3 @@
-from piazza_api import Piazza
 
 import hashlib
 import os

config/prod-root-config.yaml

@@ -28,7 +28,7 @@ chains:
     - empty.list
     - miscellanea.list
     - root-docs.list
-    - root-tutorial.list
+    # - root-tutorial.list
     # - root-forum.list
   base:
     # roles that A2rchi knows about
@@ -53,12 +53,12 @@ chains:
       AnthropicLLM:
         class: AnthropicLLM
         kwargs:
-          model_name: claude-3-opus-20240229
+          model: claude-3-opus-20240229
           temperature: 1 # not sure if 1 is best value? for now keeping consistent with prior settings.
       OpenAILLM:
         class: OpenAILLM
         kwargs:
-          model_name: gpt-4
+          model_name: gpt-4o-mini
           temperature: 1
       DumbLLM:
         class: DumbLLM

pyproject.toml

@@ -12,52 +12,60 @@ authors = [
     {name="Matthew Russo", email="mdrusso@mit.edu"},
 ]
 dependencies = [
-    "accelerate==0.23.0",
+    "accelerate",
     "backoff==2.2.1",
-    "beautifulsoup4==4.12.2",
+    "beautifulsoup4==4.12.3",
     "bitsandbytes==0.41.1",
-    "chromadb==0.4.12",
+    "chromadb==0.4.22",
     "clickhouse-connect==0.6.6",
     "coloredlogs==15.0.1",
     "duckdb==0.8.1",
-    "fastapi==0.99.1",
-    "flask==2.3.3",
-    "flask-cors==4.0.0",
+    "fastapi==0.109.0",
+    "flask==3.0.3",
+    "flask-cors==5.0.0",
     "flatbuffers==23.5.26",
     "hnswlib==0.7.0",
-    "httptools==0.6.0",
+    "httptools==0.6.1",
     "humanfriendly==10.0",
-    "langchain==0.0.268",
+    "langchain==0.3.0",
     "loguru==0.7.2",
     "lz4==4.3.2",
-    "mistune==3.0.1",
+    "mistune==3.0.2",
     "monotonic==1.6",
-    "onnxruntime==1.15.1",
-    "openai==0.27.9",
-    "overrides==7.3.1",
-    "pandas==2.1.0",
+    "openai==1.8.0",
+    "overrides==7.4.0",
     "peft==0.5.0",
     "piazza-api==0.14.0",
-    "posthog==3.0.1",
-    "psycopg2==2.9.9",
-    "pulsar-client==3.2.0",
-    "pygments==2.16.1",
+    "posthog==3.3.1",
+    "pygments==2.18.0",
     "pypdf==3.16.1",
     "python-dotenv==1.0.0",
     "python-redmine==2.4.0",
-    "regex==2023.6.3",
-    "requests==2.31.0",
-    "scipy==1.11.2",
-    "sentence-transformers==2.2.2",
+    "regex==2023.12.25",
+    "requests==2.32.3",
+    "sentence-transformers",
     "sentencepiece==0.1.99",
     "sympy==1.12",
-    "tiktoken==0.4.0",
-    "tokenizers==0.13.3",
-    "torch==2.0.1",
-    "transformers==4.33.1",
-    "uvloop==0.17.0",
-    "watchfiles==0.19.0",
+    "tiktoken==0.7.0",
+    "uvloop==0.19.0",
+    "watchfiles==0.21.0",
     "zstandard==0.21.0",
+    "langchain-anthropic==0.2.1",
+    "langchain-chroma==0.1.4",
+    "langchain-community==0.3.0",
+    "langchain-core==0.3.5",
+    "langchain-text-splitters==0.3.0",
+    "pydantic==2.9.2",
+    "pydantic-core==2.23.4",
+    "pydantic-settings==2.5.2",
+    "chroma-hnswlib==0.7.3",
+    "anthropic==0.34.2",
+    "torch",
+    "scipy",
+    "pulsar-client",
+    "psycopg2-binary",
+    "requests",
+    "pyyaml"
 ]
 
 [tool.setuptools]

test/__init__.py

@@ -0,0 +1 @@
+

test/test_chains.py

@@ -73,3 +73,5 @@ def test_chain_call_prevhistory():
     result = c1([("User", question), ("A2rchi", answer), ("User", follow_up)])
     c1.kill = True
     assert result["answer"] is not None
+
+

test_langchain.py

@@ -0,0 +1,37 @@
+from a2rchi.utils.data_manager import DataManager
+from a2rchi.chains.chain import Chain
+
+# place to fix:
+# 1.update chroma in langchain by 'pip install -U langchain-chroma'
+#   and use 'from langchain_chroma import Chroma' to replace 'from langchain.vectorstores import Chroma'
+# 2.
+
+class DiscourseAIWrapper:
+    def __init__(self):
+        self.chain = Chain()
+        self.data_manager = DataManager()
+        self.data_manager.update_vectorstore()
+
+    def __call__(self, post):
+
+        formatted_history = []
+
+        formatted_history = [['User', post]]
+
+        # form the formatted history using the post
+
+        self.data_manager.update_vectorstore()
+
+        answer = self.chain(formatted_history)
+        return answer
+
+
+# question = 'Hello, I am trying to make a box over a TF1 in the Legend. I have something like this: where fstat is a TF1 legendcell = new TLegend(0.67,0.53,0.98,0.7+0.015*(years.size()));fstat->SetLineColor(429);fstat->SetFillColor(5);fstat->SetFillStyle(1001); egendcell->AddEntry(fstat, "#bf{" + legend + "}" ,"fl"); legendcell->Draw(); The problem is that this also creates a filling in the TGraph. Is there any way to create a box on the TLegend without changing the drawing on the canvas? PS: I also tried to do fstat->SetFillStyle(0); after the drawing of the legend but this also removes the box from the TLegend My root version is 6.28/00 I really appreciate any help you can provide.'
+question = 'Hello, I am trying to make a box over a TF1 in the Legend. I have something like this: where fstat is a TF1 legendcell = new TLegend(0.67,0.53,0.98,0.7+0.015*(years.size()));fstat->SetLineColor(429);fstat->SetFillColor(5);fstat->SetFillStyle(1001); egendcell->AddEntry(fstat, "#bf{" + legend + "}" ,"fl"); legendcell->Draw(); The problem is that this also creates a filling in the TGraph. Is there any way to create a box on the TLegend without changing the drawing on the canvas? PS: I also tried to do fstat->SetFillStyle(0); after the drawing of the legend but this also removes the box from the TLegend My root version is 6.28/00 I really appreciate any help you can provide.'
+
+
+archi = DiscourseAIWrapper()
+answer = archi(question)
+print("\n\n\n")
+print(answer)
+print("\n\n\n")