PropGen

Inspiration

In enterprise sales especially with the public sector a proposal process called RFPs (Request For Proposal), RFIs (Request For Information), security questionnaires and etc. is necessary for the due diligence process in handling enterprise deals. This process is a very painful and cumbersome process that takes up high value/paid resources of searching through past documents and edits. This is a necessary but very time consuming process in enterprise sales that can take away critical talent from important tasks.

What it does

PropGen is a proposal, customer questionnaire and customer support answering tool that automates the process of analyzing all previous records and information about a company, deal or customer in order to establish context for answering questions

How we built it

PropGen is comprised of the following

• Interface layer: abstraction layer that provides 2 apis for querying and indexing
• Preprocess layer: powered by OCR to analyze documents like .doc, excel, and pdfs.
• Index layer: powered by a vector database for storing context
• LLM layer: powered by any LLM that is support by langchain such as huggingface or OpenAI

PropGen process all within docs the designated input folder and preprocesses it for embedding creation to be inserted into the vector database. Any queries can then be asked which will query a universe of context to be fed to the LLM with the question being asked to provide the appropriate answer.

Challenges we ran into

• Trial and error when finding a suitable preprocess method
• Sometimes the LLM is still prone to hallucinations and as such any questions that are very complex or out of context with the fine tuned information will not provide and suitable answers and the LLM will state that it does not know.

Accomplishments that we're proud of

This prototype exceeding expectations in results and is a good base ground for creating further implementation

What we learned

Very good introductory into understanding LLMs, LangChain and Vector databases

What's next for PropGen

• If there is good demand for this product, a chat or document editing interface can be applied on top.
• Recursive engineer can also be applied for improving the results as more proposals are provided.
• Customer context can also be applied where the universe of context can be applicable to a user record for an even better customer experience

Built With

• anaconda
• langchain
• openai
• pinecone
• python