ML/AI/LLM

@111561876209776 roadblock

Alright on it

we need to have it setup today please, i will be available after 9-10 so trusting you with this.

For website and whatsapp

They have used ByteSpeed before

Target Audience: Adult smokers (22-40 years) who want to quit smoking or reduce cigarette consumption but struggle with cravings and relapse. Type of Bot: Product Recommendation / Product Selling The bot helps users: • Understand QuitCi products • Choose the right product • Learn how the quitting process works • Get order and delivery support • Stay motivated during their quit-smoking journey Top 5 Queries The AI Must Answer Extremely Well: 1. How does QuitCi help me quit smoking? Response should explain the quitting process, Flow device, Fuds flavors, habit replacement, and craving man…

<media>

It’s not saving as a sheet

https://drive.google.com/file/d/1qcjdj3zeon5vCris7R_yEVhss_a_Xkw2/view?usp=drivesdk

share the sheet of recurring qs again that arent a ground truth

In the chat history

Based on intent == ground truth

3436 are ground truth

12.6%

I have a pie chart for that wait

check how many of these are ground truths: https://beyond-chats.netlify.app/ivfindia.com/mind-map if its a ground truth then we already have an answer and there isnt a need to cache them

complete analysis report for ivf india's past 1 year's chat

https://docs.google.com/document/d/1_JFi1jxYsADuXMqVAvg5GVc1rwLWeS0H/edit?usp=sharing&ouid=105338637044258545825&rtpof=true&sd=true

Rank,Question,Frequency 1,"i need medical advice",775 2,"my ivf cycle failed what now",771 3,"i am trying to get pregnant",769 4,"what is the cost for ivf",673 5,"how do i identify a good ivf doctor",377 6,"what is the ivf process",360 7,"why do ivf cycles fail",337 8,"i am getting ivf treatment elsewhere",326 9,"what are common ivf scams",316 10,"how can i book a zoom consultation",285

<media>

<media>

@111561876209776 dev server ka storage full ho gaya cannot login apparently

@111561876209776 Created hotfiix mr

Thank you

<media>

It should stop the bot answering and giving same/similar reply and not talk out of bounds but there are a few cases where the past generated query doesn’t match the current although they are similar that’s where the improved fuzzy matching would had worked

Working on these

quick hotfix needed for this too

quick hotfix needed for this

@155538146996345 i have gone through elefant chats, main cheez yhi pata chali ki bot is repeating messages many times + elefant ka support is not picking up chats jaha bot told ki 24-48 hrs mein call ayga which is making user frustrated

https://beyond-chats.netlify.app/elefant-687e32d809004/133952?right_drawer_tab=chat_details why low intent yaha bhi

explicitly bol rha user ki yhi order id hai, fir bot kyu nhi mann rha :( https://beyond-chats.netlify.app/elefant-687e32d809004/118418?right_drawer_tab=answer_sources&message_id=741028

bot can't say ki cancellation request sort krdenge https://beyond-chats.netlify.app/elefant-687e32d809004/123327?right_drawer_tab=answer_sources&message_id=662327

repeating same thing again and again about return and max plan https://beyond-chats.netlify.app/elefant-687e32d809004/115013?right_drawer_tab=answer_sources&message_id=721693

mujhe afsos hai..... har jagah https://beyond-chats.netlify.app/elefant-687e32d809004/115013?right_drawer_tab=answer_sources&message_id=726404

jab tak koi user puche na ki how to return toy using app, tab tak yeh mt batao, aur frustrated user ko toh mt hi batao kyunki sach mein obvious cheez hai yeh, and bahut saari chat already hogyi bot and user ke bich, agar obvious cheeze bataenge toh user ko lagega ki bot se hi baat krrha sachme https://beyond-chats.netlify.app/elefant-687e32d809004/115223?platform=whatsapp&right_drawer_tab=answer_sources&message_id=737788

iska lead score low kyu hai ?, either apan logo ko support team ke liye ek new score metrics lagana chaiye jisse team jaldi pick kre support query ya lead score hi high batana padega https://beyond-chats.netlify.app/elefant-687e32d809004/115223?platform=whatsapp&right_drawer_tab=chat_details

Got it

Thank you

access shared

@111561876209776 need access

<media>

Looking at it

we have timeline intelligence task, it will be live by 30th. It aims to track alot more of the events: https://drive.google.com/file/d/1ZU8cr8UzWBjXsp3vtTa04Ka0qkL7DTl3/view?usp=drive_link maybe some of the events can help the lead scoring system

But existing customers who need follow-up from dr malpani or needs help from elefant are a lot

I figured the current codebase does not account for existing customers

I actually was working on it

@54125194641609 Some thoughts I had regarding revamp of lead scoring: - lead scoring logic should be different for existing customers and potential customers. Major gap in our existing system is: - no differentiation of the two Some ways to mitigate this gap (in order of confidence): 1. integration with client's CRM [high confidence] 2. ask the user in the login form (only applies for website; for other platforms if any new convo is initiated then ask it once and save forever) [medium confidence] 3. try to predict via conversation [low confidence; very bad solution] Because of this majo…

@111561876209776 updated the sheet

Okay

please use notion for sharing .md docs. This doc will expire in 28 days

this list is outdated. and please do consider task updates Background-flask auto-deletion of old log files: completed Custom send-message pipeline: completed Implement repeat query detection: completed Finding & caching most-asked questions: emon just started this task Help users ask better questions: pretty sure I have removed this task since it became a sub-task of user journey mapping Train AI to reply like Dr. Malpani: this has a blocker mentioned in task updates maa-ivf callback query handling flow: added blocker in task update

@111561876209776 …

Either we need his password or a dataset which I was supposed to get from ram but it has been a blocker for now

I asked for dataset for caching, made an implementation plan for journey planning here https://www.mdshare.online/s/SxjMS0kf-c4kEtBkGoAvW

How can he give access? He won’t share his password

Background flask auto deletion I made an mr for that, custom message pipeline I was midway testing but it was already implemented, repeat query detection is live, train ai to reply like Dr malpani is pending until we get access to his email

@111561876209776 / @54125194641609 Either update Emon's tasklist w updated status and timelines, or inform me how do I oversee progress on Emon's tasks

*Product sheet: open tasks assigned to Emon (sprint 16–30 May)* •⁠ ⁠Elefant personality update — In Dev •⁠ ⁠Background-flask auto-deletion of old log files — In Dev (ETA 16/May, overdue) •⁠ ⁠Patient journey planning — Pending (ETA 30/May) •⁠ ⁠Custom send-message pipeline — Pending •⁠ ⁠Implement repeat query detection — Pending •⁠ ⁠Finding & caching most-asked questions — Pending •⁠ ⁠Help users ask better questions — Pending •⁠ ⁠Train AI to reply like Dr. Malpani (from his old emails) — Pending •⁠ ⁠beyondexams chatbot improvements — Org Verif Pending •⁠ ⁠maa-ivf callback query handlin…

I added those in why column, I’ll add a new column for this

I can

is it possible for you to add values for what the lead should have been?

I would require chats, ai answers, and if possible singular messages and their answers too

My main motive is to analyse the chats for recurring questions asked so that we can make a caching layer

what columns will you need?

@111561876209776 is there a dataset of ivf India’s last 1 years’ chats?

this has plan that me and claude made for journey plas isnning, ill implement this as soon as it is validated, this is end to end

https://www.mdshare.online/s/SxjMS0kf-c4kEtBkGoAvW

I would set up different generation logic for all, this can be a future prospect

Although it would had been better to bifurcate it would be cleaner as an org needs both support and leads gen

its just the lead generation logic in background flask that will be made custom for orgs

lead_score is a general term. it can be used in elefant’s and ivfindia’s own context

Plus as a future prospect it would be great if we had the option to type the page we want to go to as right now I was at 15th I would have to spend time to go back to 15th (click next, wait for it to load then do it 15 times) if I reloaded the page, this is not serious

Which obviously is bad for the company who’s using our service

The current lead generation does not help with that Infact when someone on elefant says they are frustrated, dissatisfied, wants a refund, wants to cancel due to poor service, will go to consumer forum, our lead analysis does not help with that as that shows consumer is not interested and gives low score

This can work for ivf India too due to a lot of users being repeated patients who are enquirin about future appointments and medications

For customer support, we should make a whole different system that would just give a support priority score high or low no medium rather than using lead analysis for it

While doing analysis on elefant I feel we should not use lead analysis at all

This sheet has my findings for lead analysis for Ivf India and elefant, Della was not suffering from wrong lead generation hence I skipped it for now

https://docs.google.com/spreadsheets/d/1k1FslLt7O-MeXxvT-CNzLCmtS7QZ5JZy6-NUq9VxajA/edit?usp=drivesdk

Okay

okay thanks. “⁠⁠finding and caching answers to the most asked questions” was missing task details, ive added them

I’ll take them up tonight

ALOT of important tasks in 1-15 and 16-30 sprint are still pending and havent been picked up - ⁠if we switched models today then how do we monitor response quality qualitatively and quantitatively? - ⁠⁠finding and caching most asked questions - ⁠LEAD_SCORE logic needs to be made org-specific - ⁠Patient journey planning

Do you have anything I can work on other than the csv

There is scope for improvement on it, ill add it for a later task then

this is already live now. Your MR already updated and merged

I forgot to tell about it, I added the new parts

Oh I updated the prompt updation doc

Repeated intent MR needs a better fuzzy matching of generated search queries

Finish the csv and do an updation to repeated intent MR

@54125194641609 tasklist for today?

fix live now

thanks. i am taking this forward, needed the fix a bit early since its affecting alot of users

User is asking if 'Hcg numeric shows according to gestational age' (3 to 4 weeks) is normal for their situation. Language will be English. The user is confused about hCG interpretation, fitting 'High Interest/Potential ICP' (Rule a). My goal is to clarify that '3 to 4 weeks' is still a qualitative assessment and that the specific numeric mIU/ml value is essential for precise interpretation. I will re-explain the calculation for menstrual age based on their transfer details and recommend the HCG AI tool again. No CTA will be included due to CTA cooldown.

but also in final message bot did not even think of this link as cta

and it thought relevancy over rules and it also got the rules wrong

working on code level enforcement

main issue is ye code enforced nahi hai, and due to ham gemini/open ai ko sidha chat milti hai and woh chat ko galat count kar raha tha iss case me

I’ll find it by 11:10

I’m checking between org services details and prompts

@54125194641609 updates?

I’ll debug

@54125194641609 can u debug why cta cooldown didn't apply? and please add a pre generation check where we exclude the cta from important_links and static_info if it was included in the prev 2 turns. the rules are already present as cta cooldowns but not enforced at code level

Why are we showing same action button again and again to the user with each response? Especially when they have already clicked on this action previously! Very frustrating

I’ll start working by 5, I’ve checked the comments I’ll add them

We could had done it on discord

also we should start tracking each bug fix somewhere. what the bug was + cause of bug + what the fix was

I’ll start today at 5, will look into them

@54125194641609 left comments on bug fix documentation: https://docs.google.com/document/d/1kaSq3jYc3XXT_TPuOJiwKbEAHVWHhYuCIQd8ddF4rM4/edit?tab=t.0#heading=h.r2g2y3pvmnv1

for both node and laravel

and rest i have tested the update and created MR

app script code

https://docs.google.com/document/d/1b-G31tUfG05kJb5u6vbjbH8-4Ru63yO-3N6HO9HtOMI/edit?usp=sharing

add Laravel logs to log fetch understand the lead analysis flow select bad examples for lead analysis (50 each for 3 orgs) and examine them one by one make 2 columns the selected chat link and why is the score not correct

In waiting room

Joining

https://meet.google.com/cna-qjzm-fhw

@111561876209776 link

Okay

lets keep it at 6:15

Is 6 fine?

@54125194641609 when can we connect?

ho gaya login abb, siddhant ne kardia

sorry for the time delay mera claude log out ho rakha hai

now the prompts hifh criteria emphasizes explicit booking intent and service selection. ravindra's chat contains payment confirmations, scheduled consult details, and procedure coordination but it also contains a lot of images, short confinmations, and staff messages. The model may be over weighting transaction and iggnoring intent leading it to default low in noisy contexts.

also the lead analysis does not see previous key info which may influence the scores as these are long conversations

service_interested_in is empty suggests the model didn’t map the conversation to services

this is for 57454 nandini's chat

<media>

this is for chat id 85194 ravindra's chat

<media>

<media>

in the chat log shared the business actions clicked is false for both and service intrested is also empty

true*

so our entirety of the lead analysis is prompt based with 2 manual parts if service intrested in matches we bump the lead analysis score by +1 and the same if business actions clicked is trye

went through the chat and log shared

Looking into this

@54125194641609 I have mailed you the relevant data to help debug this.

How is this low intent??

https://beyond-chats.netlify.app/ivfindia.com/leads?page=3&show_ai_analysis=57454 How is this low intent???

Looked upon this in perplexity and Infact scrapers/parsers playwright, beautifulsoup, cheerio read this as they ignore css and directly parses the dom by default

The offerings are dependent on what the user selects. It’s basically a single booking page for a lot of different services - controlled using various css properties. Which is not good for us, but an edge case we need to take care of

@111561876209776 can shed light on this

Did we get html then parse using bs4

How did we scrape Della?

but they are inside a display=none element

I can see that the package inclusions that were scraped are present on the page

This is related to della bug

<media>

I am getting this

Did everyone got logged out of Claude ?

2 mahine purana MR hai but I think it will help towards this, extra changes Bata Dena I’ll do them

gitlab.com/sid6i7/beyondchats-node/-/merge_requests/45

ill see to it

Did it not merge?

The repeated intent detection which I made so that AI would not give the same response repeatedly is not in the codebase

@111561876209776 just checked the prod code base

Very poor responses. What’s going on?

hallucination maybe

why did it write the plan

Thank you

I need a link for any chatbot that would trigger the node server so that logs can be made

https://dev.beyondchats.com/embed/demo/ivfindia.com

On dev server

@111561876209776 how do you create the links to test chatbot?

*not org description*

for the organisation, this new prompt would go there

jo redis se aa raha hai, there is a prompt field

this will go into the prompt field for elefant

I’ll let you know where this prompt goes

Yes yes yes

@54125194641609 is this it?

<media>

thankss

Removing all of the extra pages we are at 1230 something

Bhot bada website hai inka, cannot directly be scraped would had taken hours, around some 2000-3000 pages in total including free resources

https://www.firecrawl.dev/

used claude code with firecrawl skill for webscraping

https://docs.google.com/spreadsheets/d/1Mr9RZcdvWoYTEclOkSVXE3bsZoskALjzO_4jPZNCGNQ/edit?usp=sharing

@111561876209776 final csv with non free resource pages

https://docs.google.com/spreadsheets/d/1G5-FI-GK3p5MD1nXuR2lwUCi9QQuhQIoOn9_bO0fzbg/edit?usp=sharing

Already can share

I have this csv built

1400 ongoing

Scraped 1300 pages

@54125194641609 any progress?

Alright

final csv: title, link, description

@54125194641609 https://productschool.com/ skip "free resources" and any nested PDFs

Thank you

<media>

<media>

@54125194641609

Can you resend the screenshot

Yes it is pm2

dev should match that of prod

hope that answers your question for prod. there is log rotation present

is it pm2

@54125194641609 for node server, who handles log rotating?