Nice one! So basically lemme explain: - We did call APIs for the model, because building a model from scratch would have been difficult and expensive. But the issue with these base models were that they didn't perform well. models like the text-bison-32k were giving very low accuracy along with Gemini. It wasn't able to retrieve the right information. - Although we couldn't alter the model, we were able to change a lot of factors out of which the following worked for this problem statement. 1) Query Rephrasing: By transforming the query "What is the PPNE for 2018?" to "What is the Plant, Property, & Equipment value for 2018?", the model facilitates better matching in the documents where the acronym may not be explicitly used. The original text from SEC filings often does not directly match the query terms due to variations in terminology. By rephrasing the queries, we could bridge the gap between the query language and the language used in the documents, improving the effectiveness of text matching. 2) Dynamic Context Window Adjustment: The "dynamic adjustment" of the context window in the model refers to an iterative process where the model evaluates and adjusts the amount of text it considers generating accurate responses. Initially, the model retrieves a context that might typically suffice for answering a query. If this context, assessed by a character count (e.g., less than 100,000 characters), is deemed insufficient for a precise response, the model expands the window by including additional text or documents. This process continues until the context is sufficiently large, ensuring that the nuances and details necessary for accurate financial reporting are preserved in the model's responses, thereby adapting to the variable information density and relevance found in different sections of financial documents. These are about 2 out of the 7 points that we changed. Thanks so much for showing interest, id love to answer any more questions.
@@sohamagarwal00 so other than the simple api request and parameter changes. my main concern is promoting this half assed chatbot as ai powered when in truth it is nothing but a search and query feature for SEC filings. and saying that it still produces errors is basically stating that its a really bad search and query bot no better than a simple sql library.
Amazing team that I worked with. Loved what we achieved!
great work guys
Great Job Soham and team 🎉
All the Best. Great job
All the Best Soham 👍
Sohhaaaammmm 🎉🎉🎉🎉🎉
All the best Buddyyy ❤❤❤❤
Wishing the team success. All the best!
Great Idea.. 👍
yeah
👍👍
Interesting! Great job gentlemen!
Very nice Soham
All the best team
Great wishing Success, Success, Succes and gtoqth
All the very Best bro
All the Best,❤
Looks interesting!
Great job 🎉
Excellent innovation
Excellent work 👏🏻
Very Great.🎉
Great work
Good job!
Excellent
Good job 🎉
Nice one
All the best
Excellent 👏👏👏
All the best my dear ❤
Super
All the best 🎉
Keep it up
Superbbb
Great🎉
Great
go soham!
Impressive, how and when can we use this for the Indian markets?
❤
👏👏
Unique solution! Much needed in today's world.
Nice work with the video! Idea Seems interesting…
i agree. i am sathvik btw
That guy Agarwal in the video has a good vibe. He should play guitar or something
The video is about business management not a live concert at a cafe
All aspirants gather here
so basically rudimentary web scraping and a basic api plugin. Don't use buzzwords like ai when you dont know what it means.
If it was that simple it wouldn't have taken 6 months
@@sohamagarwal00 yeah it shouldnt have taken 6 months
Nice one! So basically lemme explain:
- We did call APIs for the model, because building a model from scratch would have been difficult and expensive. But the issue with these base models were that they didn't perform well. models like the text-bison-32k were giving very low accuracy along with Gemini. It wasn't able to retrieve the right information.
- Although we couldn't alter the model, we were able to change a lot of factors out of which the following worked for this problem statement.
1) Query Rephrasing: By transforming the query "What is the PPNE for 2018?" to "What is the Plant, Property, & Equipment value for 2018?", the model facilitates better matching in the documents where the acronym may not be explicitly used. The original text from SEC filings often does not directly match the query terms due to variations in terminology. By rephrasing the queries, we could bridge the gap between the query language and the language used in the documents, improving the effectiveness of text matching.
2) Dynamic Context Window Adjustment: The "dynamic adjustment" of the context window in the model refers to an iterative process where the model evaluates and adjusts the amount of text it considers generating accurate responses. Initially, the model retrieves a context that might typically suffice for answering a query. If this context, assessed by a character count (e.g., less than 100,000 characters), is deemed insufficient for a precise response, the model expands the window by including additional text or documents. This process continues until the context is sufficiently large, ensuring that the nuances and details necessary for accurate financial reporting are preserved in the model's responses, thereby adapting to the variable information density and relevance found in different sections of financial documents.
These are about 2 out of the 7 points that we changed. Thanks so much for showing interest, id love to answer any more questions.
@@sohamagarwal00 so other than the simple api request and parameter changes. my main concern is promoting this half assed chatbot as ai powered when in truth it is nothing but a search and query feature for SEC filings. and saying that it still produces errors is basically stating that its a really bad search and query bot no better than a simple sql library.
My man that's not how SQL works. And also it's not just parameter changes. Why so salty 😂 who hurt you, I'll help out
Dumb
Great job 👏
Great
Super