I Constructed an Auto Proposal Generator LLM and Open-Sourced It

Just lately, I haven’t been updating my open-source articles as regularly — not as a result of I’ve stopped writing, however as a result of the progress on open-source commercialization has been nice, and the limitless activity of drafting proposals has consumed my time.

As a local open-source industrial firm, WhaleOps employs largely engineers. Asking these open-source contributors to put in writing proposals wastes their growth time, and their proposals don’t at all times meet the standard normal. Not like managing in a giant firm, being a startup CEO means getting into any position the corporate wants. After each strategic assembly, I’m the primary to roll up my sleeves and deal with essentially the most pressing duties. Consequently, I haven’t had time to put in writing articles, as my restricted time is generally taken up with creating proposals that comply with the formal template type. Particularly not too long ago, with one bid after one other, I’ve discovered myself questioning my very own sanity.

The present giant fashions coudn’t clear up my drawback, so I constructed my very own.

As a tech particular person, I at all times search for instruments to resolve my issues. Naturally, I considered giant fashions, however they’ll’t totally grasp the nuances of our merchandise and infrequently produce unreliable outputs. Plus, you’re the one delivering the work, not the mannequin.

So, I made a decision to develop a software for proposal era primarily based on giant fashions utilizing Python and a appropriate giant mannequin (appropriate with ChatGPT). This software routinely generates a proposal out of your product documentation, breaking down product manuals into function factors. Primarily based on a human-created mapping of those factors to the necessities, it generates a Phrase model of the proposal and an Excel deviation desk. The mannequin can condense or broaden content material or just copy related performance as wanted.

Options of the Open-Supply Proposal Instrument

The performance of this software is straightforward, with essentially the most difficult half being the Phrase formatting (formatting in Phrase is at all times a ache). I experimented with a number of strategies to ensure it follows the right Title 1, Title 2, Title 3, physique textual content, desk, and picture codecs within the Phrase template. Staying true to the open-source spirit, I’ve uploaded the software to my private GitHub below an Apache License, so be at liberty to make use of it when you want it. Right here’s what it does:

Breaks down your product handbook right into a set of reusable element paperwork, lowering the necessity to repeatedly reference the supply doc when drafting a proposal. You can even customise particular performance (the default file identify is “Template.docx”).
Primarily based on the necessities desk crammed in by an individual, it routinely generates a proposal in a point-to-point response format, together with all headings and content material, with the right Title 1, 2, 3 formatting, and routinely organizes physique textual content, pictures, and bullet factors (the default necessities desk is “requirements_table.xlsx,” and the generated content material is in “proposal_content.docx”).
For any product necessities within the corresponding performance part, it routinely copies the product handbook content material into the point-to-point response part, retaining pictures, tables, and bullet factors. You can even rewrite the product description to swimsuit totally different proposal wants. If there’s no matching performance, the mannequin routinely generates related content material (assessment and modify as wanted).
Completes the technical necessities deviation desk by routinely filling in responses within the “requirements_table.xlsx,” with responses formatted as “Reply: Totally helps, {Mannequin-generated textual content primarily based on venture necessities}” and contains the part quantity equivalent to the proposal.

With this software, you’ll be able to shortly modify and generate proposals at will. You may get it here.

Proposal Technology Course of

Begin by working Extract_Word.py to generate your product function factors as a Phrase doc, then run Generate.py. If a function level is lacking, you’ll be able to enter “X” within the Excel file, and the mannequin will generate content material that meets the necessities. Nonetheless, I strongly suggest a handbook assessment. After era, you’ll see the proposal formatted with a desk of contents, physique textual content, pictures, tables, and bullet factors, all routinely organized.

The whole proposal is structured in line with the consumer’s necessities in sequential format, with subheadings, content material, pictures, and point-to-point responses. This takes care of all of the repetitive work.

The routinely generated deviation desk contains the whole lot, whether or not the content material was generated by the mannequin or not, together with the corresponding part numbers. You solely have to finalize the deviation desk with “&” symbols primarily based on the ultimate necessities — the mannequin has written all of it for you.

For me, this software has diminished what was once 8 hours of labor to round half-hour, and even our enterprise staff can generate the enterprise proposal sections utilizing the template. In whole, this has minimize down our time for a week-long proposal to 1-2 days, lowering the manpower required by 80%.

Use It?

First, obtain all of the code to a single listing from GitHub: Proposal Large Model (Chinese version).

Set up the Python surroundings and packages: pip set up openpyxl, docx, openai, requests, docx, python-docx.
Apply for a ChatGPT or Baidu Qianfan giant mannequin key (I used ERNIE-Velocity-8K, which is free), report the token, and place it within the related code part.
Copy your product handbook to Template.docx. Be sure you use the physique textual content, Title 1, Title 2, and Title 3 types offered; different codecs might trigger points.
Run Extract_Word.py to generate the function level doc from the product handbook (helps as much as 3 heading ranges). If the record formatting seems off, don’t fear; the ultimate format will align correctly.
Fill in Columns B and C (which can generate secondary and tertiary headings routinely) and Column G (the corresponding product handbook chapter). If a chapter is lacking, enter “X.” Notice that if there is no such thing as a corresponding chapter or an “X” is entered, the mannequin will generate content material routinely.
Assessment the “proposal_content.docx” doc and hold the chapter for which you need to begin producing the proposal. You may modify the physique textual content and heading 1, 2, and three types; simply don’t rename the types, or there could also be errors.
Regulate the parameters in Generate.py:
- API_KEY and SECRET_KEY: Baidu Cloud giant mannequin keys.
- MAX_WIDTH_CM: Most picture width; pictures bigger than this might be resized routinely.
- The prompts for producing content material have been custom-made for giant knowledge eventualities, however be at liberty to change them.
- MoreSection=1 will learn Column C to generate detailed tertiary headings (default is on).
- ReGenerateText=0 will re-generate textual content content material routinely for various proposal wants (default is off).
- DDDAnswer=1 generates the point-to-point response content material on the high of every function level (default is on).
- key_flag=1 provides the significance stage of every requirement to the proposal headings (default is on).
- last_heading_1=2 specifies the beginning chapter for the technical answer in “proposal_content.docx.”
Run Generate.py.

Abstract

With this software, you solely have to verify the product options in opposition to the proposal necessities, and a lot of the proposal content material is generated routinely. The enterprise proposal part can be generated equally, so making a 1,000-page proposal now solely takes a couple of minutes.