DSPy: The framework for programming—not prompting—language models
Python
Make langchain rag copy-able
chenmoneygithub130e3d2874e5
diff --git a/dspy/predict/langchain.py b/dspy/predict/langchain.pyindex 32e7250df..d61256995 100644--- a/dspy/predict/langchain.py+++ b/dspy/predict/langchain.py@@ -1,5 +1,6 @@import copyimport random+from functools import reducefrom langchain_core.pydantic_v1 import Extrafrom langchain_core.runnables import Runnable@@ -163,3 +164,29 @@ class LangChainModule(dspy.Module):def invoke(self, d, *args, **kwargs):return self.forward(**d).output+ def reset_copy(self):+ """Override `reset_copy()` to skip parts that cannot be deep copied."""+ if self.chain is None:+ obj = copy.deepcopy(self)+ else:+ chain = self.chain+ steps = []+ for step in chain.steps:+ if isinstance(step, LangChainPredict):+ # Only copy the LangChainPredict object.+ steps.append(copy.deepcopy(step))+ else:+ steps.append(step)++ chain_copy = reduce(lambda x, y: x | y, steps)+ # Temporarily remove the chain from `self` instance for deep copying.+ self.chain = None+ obj = copy.deepcopy(self)+ obj.chain = chain_copy+ # Put back the chain to `self` instance.+ self.chain = chain++ for param in obj.parameters():+ param.reset()++ return objdiff --git a/tests/predict/test_langchain.py b/tests/predict/test_langchain.pynew file mode 100644index 000000000..62a6665d3--- /dev/null+++ b/tests/predict/test_langchain.py@@ -0,0 +1,58 @@+import pytest+pytest.importorskip("langchain")++import os+from langchain import hub+from langchain_chroma import Chroma+from langchain_core.output_parsers import StrOutputParser+from langchain_core.runnables import RunnablePassthrough+from langchain_core.documents import Document+from langchain_openai import ChatOpenAI+from langchain_text_splitters import RecursiveCharacterTextSplitter+from langchain_community.embeddings import FakeEmbeddings+++from dspy.predict.langchain import LangChainModule, LangChainPredict++def test_copying_module():+ os.environ["OPENAI_API_KEY"] = "fake-key"+ llm = ChatOpenAI(model="gpt-4o-mini")+ docs = [Document(+ page_content="Hello, world!",+ metadata={"source": "https://example.com"}+ )]++ text_splitter = RecursiveCharacterTextSplitter(chunk_size=10, chunk_overlap=10)+ splits = text_splitter.split_documents(docs)+ vectorstore = Chroma.from_documents(documents=splits, embedding=FakeEmbeddings(size=5))++ # Retrieve and generate using the relevant snippets of the blog.+ retriever = vectorstore.as_retriever()++ prompt = hub.pull("rlm/rag-prompt")+++ def format_docs(docs):+ return "\n\n".join(doc.page_content for doc in docs)+++ rag_chain = (+ {"context": retriever | format_docs, "question": RunnablePassthrough()}+ | LangChainPredict(prompt, llm)+ | StrOutputParser()+ )+ # Now we wrap it in LangChainModule.+ rag_dspy_module = LangChainModule(rag_chain)++ copied_module = rag_dspy_module.reset_copy()+ assert len(copied_module.chain.steps) == len(rag_dspy_module.chain.steps)+ for (module, copied_module) in zip(rag_dspy_module.chain.steps, copied_module.chain.steps):+ if isinstance(module, LangChainPredict):+ # The LangChainPredict modules are deep copied.+ assert module != copied_module+ assert module.langchain_llm.model_name == copied_module.langchain_llm.model_name+ else:+ # The rest of the modules are just copied by reference.+ assert module == copied_module+ # Clean up.+ os.environ["OPENAI_API_KEY"] = None