ML/AI/LLM

when does the mind map actually gets updated? i uploaded the new csv and its showing success on training but the vectors are not loading when searched and neither is shown in recently added

Okay

just need to update the mind-map

@54125194641609 please handle this ticket asap: https://beyond-chats.netlify.app/apnipathshala.org/support?review_id=1077

Created a new migration system for node, https://gitlab.com/dev_management_beyondchats/beyondchats-node/-/merge_requests/540 Staging MR for node, description contains how to use this

Okay

FYI: scraper configs are now configurable via UI at: https://beyond-chats.netlify.app/ivfindia.com/god/crawl-setup and stored in DB

scraper config: we tell it what we dont want. and tell what paths we dont want LLMS use karke automate karna NEW LLM calls ke costs add up to tracing migration: SQL <- over script Language MR creation Background flask requirement txt and toml cleanup

Can we do aFter it then

3:30 is daily connect

Would 3;30 be fine ? My college started

please schedule a meeting anytime you’re available

@54125194641609 when can we connect today?

i also realised the missing info email updates did not have a merge request created, i cherry picked the changes in that branch onto a newer branch checked out from current laravel prod, https://gitlab.com/dev_management_beyondchats/chatgpt3_backend/-/merge_requests/1728 this is link to MR, it also includes what to do in frontend to make tracking work

also attached the test sheet in the comments of the MR

https://app.notion.com/p/Dedicated-Purchase-Intent-Classifier-3b2a4e78beb98009822fcd523c46cd99 this is documentation for product intent classification broken step by step, i have created a MR on the staging repo, please look into it

Sent an invitation

Orgs

I just added a doc in task updates for this, it has the prompt and all the changes to do for implementing this across all orbs

2 mins

username: sid6i7

@54125194641609 need access to your email-sync project on github

tl:dr not worth the trouble

research conducted on prompt cachint

https://app.notion.com/p/Prompt-Caching-3aea4e78beb9808f9b40ca70aefaf11f?source=copy_link

so for all orgs, when contact number is going to be click to call on whatsapp and click to message on whatsapp on website

the solution will apply to all

@111561876209776 this is ivf india right? and is scoped only for ivf india right?

Here docs for open claw

http://app.notion.com/p/Open-Claw-BC-scraper-3aea4e78beb9801bbabefbf2e7193186

Also dropped in Godspeed stats to compare

@111561876209776

This is the sheet and the latency comparison and why AE is faster than our node and is AE useable or not

https://app.notion.com/p/Speed-Comp-3aea4e78beb9807fbd98cc54ccdc099c?source=copy_link

Okay deta hu, but answer engine is fast because it’s a core RAG without all the added features such as suppression, query re rewrite, language detection. The llm that judges is a parallel run that just scores the answer

yr i need a side by side latency comparison of both pipelines

Mentioned in this doc

Ye Abhi kar deta hu rough steps bana Rakhe hai

answer engine is using 2.5 flash too right? then how is that faster than our system then

Nahi Maine csv bana Rakhi hai usko bas add karna hai

and this

is this done

Latency is almost same

Gemini tokenizer

new prompt: 4352 tokens (without context from vector db) ~0.001 dollars - ~0.004 dollars overall old prompt: 17920 tokens (without context from vdb) ~0.005 - ~0.01 dollars overall

how many tokens could we reduce? and whats the latency now?

On dev

The system prompt is 70% smaller now

Or multiple personas

Just test it with a persona

Prompt

what changes did we make to ivfindia on dev?

Alright, lmk when you do

not yet

@111561876209776 did you try ?

Ye cover ho jaega Isme

https://app.notion.com/p/Prompt-Compaction-3ada4e78beb9806bab37c41e80486179 research docs

sure

@111561876209776 ek baar ivf India ke bot ko apne side se test karlo on dev and let me know how it works

From here

Copy paste krdia Yaha pe

which row?

@102095734710467

I’ll now test that new prompt on dev

This was answer engine

-Test answer engine's v1 prompt on BC rubriks -test for conversations,empathy, neg statements -iterate over and over (13 times) to create a prompt that some what works

-Answer engine not that good due to structural flaws -made original godspeed docs as the benchmark -iterate over node prompt to make it shorter -made 4 version and tested all against the bennch mark ( -67%, -65%, -55%, -30%) -ported over the AE prompt to node to test for it;s accuracy cmpared to original as it was short and concise and helps us save tokens

I just filled it in daily updates wait

@54125194641609 updates?

Alright

basically add a comparison b/w answer engine and our pipeline similar to how you created for video backend vs ours. You can re-use the metrics of our pipeline from that one

Okay

@54125194641609 also, latency and token consumption analysis add kardoge kya us notion doc m?

-> Answer engine prompt is good enough to have conversations or only q/a -> Test for conversational flow -> Test conversation for empathy (sad + happy), generic counter questions is no no, focused counter questions eg asking details (relevant questions) for better answers, counter questions must not go endlessly, repeat statements are there or no -> Negative statements (eg: I cannot, I do not have, I am not sure etc even if it has a conjunction), hyperlinks productions, business actions nudge, cta rules. -> bade prompts ko chote prompts se replace -> chat history kitni acchi hai as a context

2nd one yes it will thats why we use it in async

@54125194641609 Add this as TODOs: 1. is there any reranking library in node? 2. if we use bgflask to perform the reranking and return results to node, will it add alot of network latency?

@111561876209776

http://meet.google.com/ozz-xerz-beh

updated task details

Alright

no it needs to be from scratch

Alright

if openclaw’s github page already has setup instructions then just reference those

Or I assume open claw is already setup

Will this require also how to setup open claw?

add whatever scripts/helper functions you had created

till then can you work on this task

Okay! Is three going to be fine ?

Also this

.

please schedule a meet with me to discuss about it

.

Yes

@54125194641609 any progress on answer engine evaluation task?

but not the ground truths, business actions

okayy no worries, ive set it up

Time lagega thoda I just left home for a doctor checkup I don’t think I’ll make it back by 6

@54125194641609 org setup kardo dashboard pe

kardia

hex code du?

https://app.notion.com/p/Org-Setup-Task-List-3aaa4e78beb9809297f4e053a2fb2682?source=copy_link

expected you to fill these in too

make this edittable link quickly

@111561876209776 '

task list for setup

https://gossamer-brake-54e.notion.site/Org-Setup-Task-List-3aaa4e78beb9809297f4e053a2fb2682?source=copy_link

blogs

https://docs.google.com/spreadsheets/d/11_abLVJhsIU8uHJEC4_uEsCp2UChuMKtGtJGj0zq7o4/edit?usp=sharing

main csv

https://docs.google.com/spreadsheets/d/1tGMA3Wyy9KzyA7PFO_U_A2LG6A8M01nkb2TiSXBQ_sE/edit?usp=sharing

alright

if it wont take long then blogs ki alag csv prep karlo

so kb will have twice the medical info for some surgeries

yeah, but they may collide with the surgeries page as that has similar medical info too

do they all follow the same pattern?

how easy are blogs to scrape?

i went through the website, we dont need to scrape gallery, patients we served as that is totally image based and our blogs are mainly just medical knowledge

working on this on urgent basis

understood

FYI: you dont need to create the org yourself

@54125194641609 https://www.siddhivinayakhospitals.org/

@111561876209776 I forgot to mention, answer engine uses flash rank and you mentioned flash rank had memory leaks, so be cautious while using it, its setup in dev server but quadrant is not, so to test just setup qadrant and it will be running

covers my findings of what worked what did not and what was caused that it did not work

report for answer engine test on bugsheet

https://gossamer-brake-54e.notion.site/Answer-Engine-Review-3a8a4e78beb980e4b1b7de4d574a9f80?source=copy_link

@111561876209776

also i created this codebase reference because i was shifting to codex for next few tasks, this makes it consume less tokens and have context, can be useful hence shared

https://app.notion.com/p/BC-NODE-3a6a4e78beb980bb99f4f751c900dadc?source=copy_link

this has proposed workflow we can set up based on the ivf india's prompt

https://app.notion.com/p/IVF-india-Workflow-builds-3a6a4e78beb980b7a91de1aac62a20e9?source=copy_link

this links to docs and the MR

https://docs.google.com/document/d/1T0jC4qgCWhHDg4dyTzxTDYhU7LoAa8lrZFj0mVJCcCY/edit?usp=sharing

- how much storage would redis take per 1000 chats based on no. of queries cached - Share docs on how to create the B-Tree on redis with Q/A - take any organisation and create different scenarios based on the prompt, we are referencing our existing prompt -multiple states -multiple dfd states starts with level 0 ------ After Feedback -------- -end goal <- we are able to understand that our ivf indias prompt considers -> user aaya user ne history batai but who additional level hai and usme empathy show har level me kar skte hai, we will not consider empathy as a scenario -> perspective of u…

Sharing

and caching

@54125194641609 share doc for vector db optimisations

@111561876209776

http://meet.google.com/tqz-bbkx-sxe

yes

1:30 fine?

schedule anytime before 6 or after 8

@111561876209776 when would you be free today? Let’s have a meet then

lets connect tomorrow and discuss the results, cuz after this only one thing remains that is HNSW and i do not want to implement it yet as it is memory intensive and always on the memory for better speeds

https://docs.google.com/spreadsheets/d/1FNVEYg5InqUcGXD7ikjpTwsaNhfhGiKHcENZOhc_5ew/edit?usp=sharing

i will share the exact figures but this is cached embedding figures vs non cached embeding figures

on my local db and cache + b trees are very fast

and b trees

i implemented a cache

whats the status on reducing answer latency/vector search latency by half?

replied to the support ticket with MR link and the sheet link

dev_management_beyondchats

where do i make a MR?

shone node branch kaha gaya

@111561876209776

Okay

I thought the eta was tomorrow eod hence had a slower pace

i will take it forward then. please move to other tasks

Working on it would do it by tonight

update?

Okay okay got them as support tickets

check your email’s spam or inbox (emonganguly2005@gmail.com)

I cannot see on my sheet

Alright

i have assigned 2 realtime requirements related to dr priyanka to you, eta is tomorrow EOD. Please do check

lets connect today whenever available

@54125194641609 updates? what have u been working on?

I already fixed them in the MR

is this not done?

Prompt toh prompt engineering section se hoga na? Waha toh the prompt gets updated not overidden

Okay

I created MR for this on prod

also I have assigned nithin to change dr priyanka’s persona from speaking in 3rd person to 1st person. just a FYI. I had also changed the prompt before but i guess it got overridden because i didnt communicate the change

is this live?

these look great

@54125194641609 need you to create tickets on dr priyanka’s support page for all the issues found please

i have created a MR on prod for this

this 4th column

https://docs.google.com/document/d/1d_8f7Y894H7qzux3o3TbRACBpx5DijAs6DbMu9NKLRo/edit?usp=sharing

which case?

@111561876209776

can you check the results, i'll make lunch in the meanwhile

but iss specific case ko fix karne ke chakkar mein prompt regress kar raha tha

Ye sab sach m ho rha h ya reel k liye

101 mein there was 1 comment which was classified as a troll which could had been replied that was

i have highlighted with yellow the social issues

this has 100 comments from the last 24 hr on dr priyanka

and social issues jisme raise ho rahi hai woh dm karne ko bol raha hai

ek baar results check karlo it is deflecting troll comments better now

https://docs.google.com/spreadsheets/d/1NxY8mkjq5hCTenptAQcrWCb0_1EOVtIOVhwRPlBNbQ8/edit?usp=sharing

this got classified as emoji

Thanks a ton! ❤️

Pada kr bacho dimaag kharab ho jataa mam 👏👏👏👏

emoji RE are failing

what else?

these are not the only bugs you shared

i am going through threads ka csv

the bot has reached its limit for today. It will no longer reply. Limit will reset tomorrow so lets fix its behavior till then

complaints related to social issues that are relevant to dr priyanka -> dm troll comments -> irrelevant and the but shouldnt just agree with the user comments. Too many trolls on insta, if the bot agrees with any of the troll comments then it can become a huge problem

is deflecting to dm the correct thing here? we only reply to clinic questions in dms so if someone does dm they wont actually get any productive outcome from this

This was the general template I was following for comments