How to write similar document using examples?

[qingtian1771]

Following is concreate requirements:

1. User or System can provide some example ducuments about a topic, for example: divorce agreement.
2. The writting assistant(AI) using these documents to summerize some key points.
3. The writting assistant(AI) use the summerized key points to ask user questions to collect useful information.
4. After the collection, The writting assistant(AI) will generate a document, for example a devorce agreement.

Maybe user can do this through a chatbot like following:
User: I want to write a divorce agreement.
AI: Wait a moment, I will ask you some questions, could you please answer following questions:
.....

Thanks a lot!

[dosubot[bot]]

🤖

To develop a writing assistant using LangChain that can generate a document like a divorce agreement by summarizing key points from example documents, asking the user questions based on these points, and collecting the user's responses, you can follow these steps:

1. Summarize Key Points from Example Documents: You can use a language model to summarize key points from example divorce agreements. You can create a new PromptTemplate that instructs the AI to summarize the document. Here's an example of how you might modify the code:

template = """You are an AI trained to summarize key points from documents. 
Given the following document, please generate a summary based on that document.

Example Format:
<Begin Document>
...
<End Document>
SUMMARY: summary here

The summary should be concise and be based explicitly on information in the document. Begin!

<Begin Document>
{doc}
<End Document>"""
PROMPT = PromptTemplate(
    input_variables=["doc"],
    template=template,
)

2. Ask User Questions Based on These Points: You can use the QUESTION_PROMPT template to ask the user questions based on the summarized key points. The QUESTION_PROMPT template takes a context and a question as input variables. The context would be the summarized key points and the question would be the specific question you want to ask the user based on these points.

3. Collect User's Responses: You can collect the user's responses to these questions. You can use the COMBINE_PROMPT template to combine the user's responses with the summarized key points to generate the final document. The COMBINE_PROMPT template takes the summaries (user's responses) and a question as input variables. The question would be the specific question you asked the user and the summaries would be the user's responses to these questions.

4. Generate the Final Document: You can use the collected responses and the summarized key points to generate the final document. You can use the COMBINE_PROMPT template for this purpose.

For the chatbot interaction, you can override the ConversationalChatAgent to ask for context from the user. You can refer to the solution suggested in the Asking for user input as tool for agents issue.

Please note that this is a simplified example and the actual implementation would involve more complex logic to handle multiple questions and responses, and to generate a coherent final document.

Sources

• Asking for user input as tool for agents
• libs/langchain/langchain/chains/qa_with_sources/map_reduce_prompt.py
• libs/langchain/langchain/evaluation/qa/generate_prompt.py
• libs/langchain/langchain/indexes/prompts/knowledge_triplet_extraction.py

---

This response is meant to be useful and save you time. 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.

If you want to continue the conversation, start your reply with @dosu-bot.

[dosubot[bot]]

🤖

Hello,

Based on your requirements, you can use LangChain to develop a writing assistant that can generate a document like a divorce agreement by summarizing key points from example documents, asking the user questions based on these points, and collecting the user's responses. Here's a step-by-step guide on how you can achieve this:

1. Summarize Key Points from Example Documents: You can use a language model to summarize key points from example divorce agreements. You can use the EXAMPLE_PROMPT template to feed the content of the example documents to the language model and get the summarized key points.

from langchain.prompts import PromptTemplate

# Summarize key points from example documents
example_prompt = PromptTemplate(
    template="Content: {page_content}\nSource: {source}",
    input_variables=["page_content", "source"],
)
summary = example_prompt.fill({"page_content": example_document_content, "source": "example_document_source"})

2. Ask User Questions Based on These Points: You can use the QUESTION_PROMPT template to ask the user questions based on the summarized key points. The QUESTION_PROMPT template takes a context and a question as input variables. The context would be the summarized key points and the question would be the specific question you want to ask the user based on these points.

# Ask user questions based on these points
question_prompt = PromptTemplate(
    template=question_prompt_template, input_variables=["context", "question"]
)
question = question_prompt.fill({"context": summary, "question": "What is your spouse's name?"})

3. Collect User's Responses: You can collect the user's responses to these questions. You can use the COMBINE_PROMPT template to combine the user's responses with the summarized key points to generate the final document. The COMBINE_PROMPT template takes the summaries (user's responses) and a question as input variables. The question would be the specific question you asked the user and the summaries would be the user's responses to these questions.

# Collect user's responses
user_response = input(question)

# Generate the final document
combine_prompt = PromptTemplate(
    template=combine_prompt_template, input_variables=["summaries", "question"]
)
final_document = combine_prompt.fill({"summaries": user_response, "question": "What is your spouse's name?"})

Please note that this is a simplified example and the actual implementation would involve more complex logic to handle multiple questions and responses, and to generate a coherent final document.

For the chatbot-like interaction, you can override the ConversationalChatAgent to ask for context from the user. You can find more details about this in the Asking for user input as tool for agents issue.

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

Best,
Dosu

Sources

• Asking for user input as tool for agents
• libs/langchain/langchain/chains/qa_with_sources/map_reduce_prompt.py
• libs/langchain/langchain/evaluation/qa/generate_prompt.py
• libs/langchain/langchain/indexes/prompts/knowledge_triplet_extraction.py

---

This response is meant to be useful and save you time. 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.

If you want to continue the conversation, start your reply with @dosu-bot.

[qingtian1771]

Are these anwsers provided by a robot?

[devstein]

Yes! Dosu is a Github app

[qingtian1771]

I think Dosu's answers are not quite right.
this situation is a kind of active AI and not a passive AI. Human can do this quite well, human can read some example documents and know what should ask, and can ask user some questions which will be used to collect information for document writting.
And how AI can finish this process?

I think at least it needs three steps:

1. AI will generate questions according the example documents. (maybe this step is very hard!!!)
2. use generated questions to generate a prompt, and AI use this prompt to communicate with user to collect information, the information is the chat history.
3. AI use chat history maybe other prompt, to generate a document.

[devstein]

I agree. There is a lot of room for improvement. Dosu's answers aren't meant to be a replacement for human responses but rather a tool for the community to use

[qingtian1771]

I think this problem is interesting, anyone else interested?

[JoffreyLemery]

I am ! I Just start to think about how to do it. If you want to colabore, I'm in !

[josefinarcasanova]

Hi!

This is a very interesting topic. We're working on a document generation use case. We aim to offer a chat interface to users can modify the document in real time, after an initial draft is generated.

We're struggling finding the best way to do this and would love to hear ideas and real-life experiences.

Here's a diagram of what we're thinking so far:

Basically:

• We generate a complete doc draft based on a template (similar document) the user provides.
• User requests one or more operations (adding, modifying or deleting existing content).
• An LLM detects the operation and its specifications, giving a JSON output with the parameters.
• Specific prompt for performing the operation is invoked.
• New doc version is displayed so user can see results.
• Endless loop until user is satisfied.

We're in validation stage and would appreciate any input.

[purpose-impact]

@qingtian1771 were you able to find the solution to your usecase. I am looking for something similar. It will be good know the full solution.