Scanned PDF Langchain Hybrid Search

[@ramareddymathsacademy]

Great Going. Keep up good work. Let me introduce to my students. It's really benefits their career growth. Thanks

[@chiragjain157]

I tried this method , It worked for use case I was working. Thank You

[@AIwithTarun]

Glad it helped

[@bydmatt4499]

Is insane!!! amazing video man! when will you get the entire course? I can't wait to check it out!

[@vladimirolezka3482]

Great video.❤ Bhai, could you please make a video using Qdrant as vectorstore instead of ChromaDB. Thanks 🙏

[@AIwithTarun]

Noted.

[@RashomonAI]

Quality video as always

[@AIwithTarun]

Glad you enjoyed it

[@int8float64]

good tutorial! I just have a small problem plz help the llm prints generates the whole template as well as the answer together. also the answer generated is not complete, it stops at the middle. Any help?

[@AIwithTarun]

Use model_kwargs and define max_new_tokens to be 512 to avoid answer generation in middle. Also regarding the answer printing the entire context, it's seems there is kind of changes in prompt template. Initially following template fixes this. But from the latest update it seems you can directly pass the prompt to the chain.

[@PsPappu-j3o]

thanks for great video and efforts , can you please help me how to print only assistant answers rather than whole information in the output.

[@AIwithTarun]

Usually answer should not contain the context information if you follow the prompt template as per the LLM. But if you followed it in right way, use regex Or Python List to get response after

[@CryptoMaN_Rahul]

underrated

[@naveenpandey9016]

can we implement this for JSON which contains large amount of data

[@AIwithTarun]

Yes you can. Refer: python.langchain.com/docs/modules/data_connection/document_loaders/json

[@naveenpandey9016]

​@@AIwithTarunwhich LLM model is good for generating summaries from large data. Will mistral perform well?

[@Hellow_._]

I dont understand why we have used keyword retriever here, bcz we dont have multiple documents here that you mentioned in the slides i.e (smartphone, VR, AI). cant we use only vector retriever here?

[@AIwithTarun]

The intention here is to show the implementation of Ensemble Retriever, once we complete the RAG video series, we have 2-3 end-to-end project implementation. During that time we will have multiple documents. I guess I will mention this in next video, thanks for pointing this concern, really appreciate it.

[@Hellow_._]

@@AIwithTarun thank you for replying. Yesterday Interview asked me one Question - suppose we are building application for publisher house and they have data in magazines, podcasts and in video format. And if user ask query then it should give answer from the documents. So i guess in this keyword retriever will work, right?.

[@AIwithTarun]

@@Hellow_._ In this case you also need to check how you index the data(create embeddings for the chunks) and what metadata to include. With the help of metadata, you can then use a keyword retriever. But again, I would prefer testing it with Hybrid Search.

[@Hellow_._]

@@AIwithTarun okay. Thanks. Appreciated 🙏

[@tirthbhatt3340]

Hi Tarun, i have multiple json files and i want to create q and a out of it also want to fine tune it with gpt4 could you please help me which approach and tools i chose for that?

[@AIwithTarun]

Use TRL library along with Open Source LLMs

[@nitinsharma4922]

Sir please make a video on proper roadmap for ML and AI. If possible then start a program from scratch.

[@AIwithTarun]

That's nice suggestion. If channel hits 3K subscribers I will start AI/ML video series from scratch for free based on industry standards

[@scitechtalktv9742]

What to change in the code if my Scanned pdf is in the German language? Certainly the embedding model has to change? And perhaps something else ? Also I need a solution for when I have a whole set of PDF’s in English and German, and also Scanned and normal digital PDF’s amongst them. What to do in that case?

[@AIwithTarun]

You need to check if there is any Bilingual embedding model for English + German. If not use any Multilingual. Also you need to update the System Prompt. Unstructured based loading helps with both scanned and normal pdf data