ConversationalRetrievalQAChain

[nima-cp]

I can't use any memory to give the bot the chat_history. The error is:

Error: input values have 2 keys, you must specify an input key or pass only 1 key as input

Also, it seems Langchain is updating and the schema is changing as well. here is my code:

const memory = new BufferWindowMemory({ k: 1 });
const chain = ConversationalRetrievalQAChain.fromLLM(
    model,
    vectorStoreRetriever,
    {
      memory: memory,
      qaTemplate: QA_PROMPT,
      questionGeneratorTemplate: CONDENSE_PROMPT,
      returnSourceDocuments: true, 
    },
  );

[dosubot[bot]]

Answer generated by a 🤖

Answer

Hello,

The error you're encountering, "input values have 2 keys, you must specify an input key or pass only 1 key as input", is likely due to the fact that the ConversationalRetrievalQAChain class expects two keys in the input values: question and chat_history. These keys are defined in the inputKeys getter method:

get inputKeys() {
  return [this.inputKey, this.chatHistoryKey];
}

If you're not providing these keys when calling the _call method, you'll encounter this error. To resolve this, ensure that you're passing an object with these keys to the _call method. For example:

const chainValues = {
  question: "Your question here",
  chat_history: "Your chat history here"
};

const result = await conversationalRetrievalQAChain._call(chainValues);

As for adapting to the schema changes in the latest LangChain update, the ConversationalRetrievalQAChain class has a fromLLM static method that creates an instance of the class from a language model and a retriever, with various options. This method could be useful if the schema changes involve the way instances of the class are created.

In your code, the BufferMemory class is used to handle memory. It is initialized with an object that includes memoryKey, inputKey, outputKey, and returnMessages properties. The memoryKey is set to "chat_history", the inputKey is set to "question", and the outputKey is set to "text". This means that when you call the chain.call method, you should pass an object that includes a "question" key. The "chat_history" key is handled by the BufferMemory class, and the "text" key is used for the final conversational output of the chain.

If the schema changes involve the way chat history is passed, you can also pass the chat history in explicitly by omitting the memory option and supplying a chat_history string or array of HumanMessages and AIMessages directly into the chain.call method.

If you want to further change the chain's behavior, you can change the prompts for both the underlying question generation chain and the QA chain. You can pass an alternative prompt for the question generation chain that also returns parts of the chat history relevant to the answer, allowing the QA chain to answer meta questions with the additional context.

In your code, it's also shown that you can use the returnSourceDocuments option in the fromLLM method to return the source documents used in the conversation. This could be useful if you want to track the sources of the information used in the conversation.

Additionally, you can override the default prompts used in the chain by passing a qaTemplate or qaChainOptions with a custom prompt to the fromLLM method. This allows you to customize the way the chain generates questions and answers.

Finally, you can use the chain.call method without passing a chat_history key if you're using a BufferMemory instance with the memoryKey set to "chat_history". In this case, the BufferMemory instance will handle the chat history for you.

I hope this helps! If you have any other questions or need further clarification, feel free to ask.

Sources

• langchain/src/chains/conversational_retrieval_chain.ts
• examples/src/guides/expression_language/cookbook_conversational_retrieval.ts
• docs/docs/modules/chains/index_related_chains/conversational_retrieval.mdx
• examples/src/chains/conversational_qa_built_in_memory.ts
• langchain/src/chains/tests/conversational_retrieval_chain.int.test.ts

---

This response is meant to be useful, save you time, and share context. It is not meant to be a precise solution, but rather a starting point for your own research.

Help me be more useful! Please leave a 👍 if this is helpful and 👎 if it is irrelevant.

[Drahokma]

You have to add memory key chat_history to the buffer memory:

const bufferMemory = new BufferMemory({
returnMessages: true,
memoryKey: "chat_history",
chatHistory: chatHistory,
});

[nima-cp]

Thank you for your answer. Actually I tried this and lots of other ways but eventually I switched to python and wrote a microservice for the whole langchain job.