RAG fail with TypeError: argument 'text': 'HumanMessage' object cannot be converted to 'PyString'

[ShahriyarR]

[This question is also asked in Discord] Tagging @fjsj for visibility

Hello dear all,
I would like to ask if you have already played around with the RAG feature of the library.
I am currently getting a fail as:

File "/home/shako/REPOS/EpicLaunchX/core/.venv/lib/python3.11/site-packages/langchain_openai/embeddings/base.py", line 441, in _tokenize
    token = encoding.encode_ordinary(text)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/shako/REPOS/EpicLaunchX/core/.venv/lib/python3.11/site-packages/tiktoken/core.py", line 70, in encode_ordinary
    return self._core_bpe.encode_ordinary(text)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
TypeError: argument 'text': 'HumanMessage' object cannot be converted to 'PyString'

Calling assistant as:

# Calling assistant
pattern_ai_assistant = CodePatternAIAssistant(user=request.user)
    output = pattern_ai_assistant.run(
        f"I need a match pattern for given task with uuid {task.uuid} and code implementation section as {task.code_implementation}."
        f"Please generate the match pattern for the given task code implementation",
        thread_id=thread.id,
    )

Assistant itself:

class CodePatternAIAssistant(AIAssistant):
    id = "code_patterns_assistant"
    name = "Code Patterns Assistant"
    instructions = (
        "You are a project assistant, who creates the code pattern to match the code implementation of the given task and saves it to the database. "
        "Based on the code implementation please generate patterns. "
        "Use the following pieces of retrieved context from Pattern Syntax doc to generate the code pattern. "
    )
    model = OPENAI_MODEL
    has_rag = True
    _user: User

    def persist(self):
        path = Path(__file__).parent
        file_path = path / "data/pattern-syntax.md"
        persist_dir = path / "data"
        loader = TextLoader(str(file_path))
        documents = loader.load()
        text_splitter = MarkdownTextSplitter()
        split_documents = text_splitter.split_documents(documents)
        embeddings = OpenAIEmbeddings()
        vector_store = Chroma.from_documents(
            documents=split_documents,
            embedding=embeddings,
            persist_directory=str(persist_dir),
            collection_name="patterns",
        )
        vector_store.persist()

    def get_retriever(self) -> BaseRetriever:
        # NOTE: on a production application, you should persist or cache the retriever,
        # updating it only when documents change.
        path = Path(__file__).parent
        persist_dir = path / "data"
        self.persist()
        embeddings = OpenAIEmbeddings()
        vector_store = Chroma(
            collection_name="patterns", embedding_function=embeddings, persist_directory=str(persist_dir)
        )
        return vector_store.as_retriever(
            search_type="similarity_score_threshold", search_kwargs={"score_threshold": 0.5}
        )

As I understand it tries to tokenize the user message and fails.

content="I need a code match pattern for the given task with uuid 35f72640-3aa7-447d-bddb-e2b6cb8592c9 and code implementation section as class Operands:\n    def __init__(self, first_operand: int, second_operand: int):\n        self.first_operand = first_operand\n        self.second_operand = second_operand\n\n# Ensure type safety by checking types\n    @property\n    def first_operand(self):\n        return self._first_operand\n\n    @first_operand.setter\n    def first_operand(self, value: int):\n        if not isinstance(value, int):\n            raise ValueError('first_operand must be an integer')\n        self._first_operand = value\n\n    @property\n    def second_operand(self):\n        return self._second_operand\n\n    @second_operand.setter\n    def second_operand(self, value: int):\n        if not isinstance(value, int):\n            raise ValueError('second_operand must be an integer')\n        self._second_operand = value\n.Please generate the code match pattern for the given task code implementation" additional_kwargs={} response_metadata={} id='750'

[fjsj]

I will test this manually and reach back to you soon!

[ShahriyarR]

Any updates? :)

[pamella]

Hi @ShahriyarR, I’m reviewing this as well and will share an update later today.

[pamella]

@ShahriyarR We've just released version 0.1.1, which includes the fix from PR #187. Could you try upgrading to the latest version and let us know if the issue is resolved? Thanks!

[ShahriyarR]

Wow :) Thanks ) I will check it this evening and will update the thread)

[ShahriyarR]

Okay, I do confirm that there is no such failure anymore. Thank you very much for the fix, going to continue with my experiments :)

[pamella]

Awesome, that's good news! 🎉