while building AI applications I noticed a perpetual need for a really good summarization engine for applications to work well.
Voice transcripts are a common use case. I've a long text transcript that I love to throw at an LLM to give me insights for:
1. generating summary of insights relevant for learning/researching/sales prospecting about X from the Zoom transcript
2. generate question and answer pairs using YouTube videos as a synthetic data finetuning data source
3. Synthesising most important facts of multi long documents in Perplexity-type of AI search
After going down the rabbit hole of past and present solutions, I found most projects leverage long context length LLMs, which I wasn't too satisfied with. Studies have also shown this is not a long term solution since such LLMs seem unable to extract facts from the early sections of long text.
Inspired by code agents, I decided to try an agent-based approach to summarization. Like code agents who significantly improve when used in an agentic approach instead of a zero-shot way, I implemented a hierarchical approach to generating detailed summaries with long text documents like research papers.
boy, I was blown away by the results. You can view them in the Google Docs I attached.
What do you think of these summaries? I would love the community's feedback, especially on summarization techniques that have worked well from your experience that don't require significant set up.
Disclaimer: I did not fact check the figures in the summaries so they might be subject to some hallucination
will the community be keen to see more of this project?
while building AI applications I noticed a perpetual need for a really good summarization engine for applications to work well.
Voice transcripts are a common use case. I've a long text transcript that I love to throw at an LLM to give me insights for:
1. generating summary of insights relevant for learning/researching/sales prospecting about X from the Zoom transcript
2. generate question and answer pairs using YouTube videos as a synthetic data finetuning data source
3. Synthesising most important facts of multi long documents in Perplexity-type of AI search
After going down the rabbit hole of past and present solutions, I found most projects leverage long context length LLMs, which I wasn't too satisfied with. Studies have also shown this is not a long term solution since such LLMs seem unable to extract facts from the early sections of long text.
Inspired by code agents, I decided to try an agent-based approach to summarization. Like code agents who significantly improve when used in an agentic approach instead of a zero-shot way, I implemented a hierarchical approach to generating detailed summaries with long text documents like research papers.
boy, I was blown away by the results. You can view them in the Google Docs I attached.
What do you think of these summaries? I would love the community's feedback, especially on summarization techniques that have worked well from your experience that don't require significant set up.
Disclaimer: I did not fact check the figures in the summaries so they might be subject to some hallucination
will the community be keen to see more of this project?