Langchain + Qdrant Cloud | Pinecone FREE Alternative (20GB) | Tutorial

[@alejandro_ao]

💬 Join the Discord Help Server: link.alejandro-ao.com/981ypA ❤ Buy me a coffee (thanks): link.alejandro-ao.com/YR8Fkw ✉ Join the mail list: link.alejandro-ao.com/o6TJUl

[@AIdevel]

Hi I lost my embeddings on qdrant tell how to save the embeddings on qdrant to persistant ??

[@pulkitarora6605]

@@AIdevel .recreate_collection() delete all the embeddings and recreate the collection. Comment this code once collection is created

[@ilsms]

Very very very great video we are waiting for more such videos

[@sulyatt]

Thank you very much for all your hard work Alejandro, your videos have been absolutely indispensable for creating applications using the power of OpenAI. I am really looking forward to what further tutorials you come up with that can allow us to use as a platform to make magical new tools!

[@alejandro_ao]

this means a lot! i will be posting more videos as soon as i come back from vacation :)

[@pdamartin4203]

Another great video tutorial, which actually answers my question from your previous video on how to avoid recreating the vector store for the documents whenever the app is launched, especially where the documents are fixed in the application (you do not allow users to upload documents). Muchas gracias.

[@RamiH-q8c]

Thanks for the great content

[@alejandro_ao]

No problem! Thanks for the tip, you rock 🤘

[@RamiH-q8c]

@@alejandro_ao Just found out u have discord. thats very nice of u

[@alejandro_ao]

Qdrant is an open-source vector database. Their free cluster comes with: - 20GB of storage - 1GB of RAM - 0.5 vCPU What other vector databases should I cover? 👇

[@RealEstate3D]

As already said: It would be interesting to save vector embeddings in relational databases and then when these embeddings are used at a later point in time to import them in a vector database. And how to convert embeddings into another “dimensionalaties” removing complexity.

[@alejandro_ao]

@RealEstate3D why not simply use a vector database once you’re at it? if you need several collections, you can always store the collection information in a more conventional db like mongodb

[@RealEstate3D]

@@alejandro_ao that’s the point. In my project for example I have known problems with Pinecone as they seem to have a problem with OAuth Users. This can happen. We know that s**t happens. That’s ok. But if you have 4000 law pages with metadata per paragraph and a provider can’t deliver I’d like to switch the embeddings. And here we come to the point. Before saving the chunks we should also hold it in an intermediate DB. But the way to import embeddings with metadata into other db’s isn’t exactly easy documented.

[@RealEstate3D]

@@alejandro_ao I switched from Pinecone to pgvector for gathering app key data. And I can only recommend AI app developers to use some kind of dockerized alternatives and measure the size of needed space after inserting 2000 pages, measure time of similarity search, upload further 2000 pages and redo. Beside let run a cron to snapshot system resources during these tests.

[@fozantalat4509]

@alejandro_ao can we save the embeddings locally (persistent FAISS, or chromadb), if so can please make a video on it.

[@Axel__Avila]

Great video! Glad you're back!

[@alejandro_ao]

glad to be back!

[@kevinl.9657]

Not really related to langchain but I just gotta ask, how do you zoom in at 13:51? I don't know how to zoom in/out in a browser using mouse scroll on a Mac.

[@alejandro_ao]

i just pinched the trackpad hehe it should work on chrome

[@kevinl.9657]

@@alejandro_ao haha. Figured out after I posted my comment. Looks like trackpad is the only solution.

[@NiravNandu]

Thanks for the in depth step by step process to setup qdrant

[@alejandro_ao]

No worries!

[@thirdeyeinthemaking7327]

Thank you for taking time to provide this useful information, Alejandro!!

[@alejandro_ao]

it’s my pleasure!!

[@RealEstate3D]

It would be interesting to save vector embeddings in relational databases and then when these embeddings are used at a later point in time to import them in a vector database. Or how to convert embeddings into another “dimensionalaties” removing complexity. Edit: I am using pgvector so pg dumping and restoring does the trick, but reformatting for Pinecone, Chroma and Faiss?

[@chougaghil]

I do not understand your problem If you know of to create embeddings, you know how to adapt for the db A, B or C The base of embedding is just a vector (an array of digits) Extract of a working code with python; from sentence_transformers import SentenceTransformer model = SentenceTransformer('all-MiniLM-L6-v2') embeddings = model.encode(documents) #docuement are an array for your texts #to save your embedding, just follow the api doc of your target vector db #for pinecone: pinecone.create_index(index_name, dimension=384, metric="euclidean") #dimension=384 because it is the # of dimensions of my 'all-MiniLM' model index = pinecone.Index(index_name) array_vectors=[ ] for i in range(0, len(embeddings) ): embedding=embeddings[i].tolist() pinecone_id="embedding_name"+"-"+str(i) array_vectors.append( (pinecone_id, embedding) ) index.upsert(array_vectors)

[@dawn_of_Artificial_Intellect]

great video you went over everything in great detail but one thing you did say why did you select "stuff" as a chain type or maybe I missed it?

[@kumarrohit3021]

If I upload more document and run the code, it process the existing text again and even though it was already converted to vector and uploaded to qdrant. How Can I avoid these duplicated

[@kadirgultekin7984]

the video is great, but why are our old questions not on the screen and the avatars we created for user and client disappeared?

[@韭菜盒子-u8p]

thank you very much. what if i want store csv files to vector database?

[@nickfrische8262]

You would have to change the get_chunk function, or just make a new function depending on how you would want to chunk your CSV. If it were each row, make the function so that each row combines all of the columns of the row and then makes an array of the rows. This solution might not be the best because you might need to manipulate the data so that each cell had the column heading before it but it would work along with the code in the video. Chat GPT could also help with your question

[@philiphess6859]

@@nickfrische8262🔒😈🔑

[@abhaypkyek]

Actually I switched to Qdrant after watching this video on my on premise server. Only issue in facing is duplicatation of content. How to handle that?

[@arturgomes1654]

Great content and demonstration, thank you so much!

[@mehulpathak2245]

"Connection was forcibly closed by a peer" -- I'm getting this error when I'm using thunder client and adding curl link and api key as header. What to do?

[@fbravoc9748]

Amazing video!!! Thanks for all the shared knowledge. I was wondering if you could tell me if ii si possible to update the information inside the vector database every 30 minutes, so that it contains updated information that users can query. Do you think this is possible in Qdrant?? Thanks

[@miguelalerts4318]

Very cool! Do you have a way to get the size of the embeddings created to understand if those are able to be stored in the current collections box?

[@alejandro_ao]

you can check in the documentation of your embedding model. they always mention that. alternatively, you can check the embeddings leaderboard on huggingface, it also lists their dimensionality alongside theirs ranking

[@tonimigliato2350]

Great tutorial! Thanks for sharing. What if I want to insert some template, with some instructions like "You're a helpful assistant" and "if you don't know the answer say you don't know"? I think I would probably use LangChain PromptTemplate for that but, where would I insert that code?

[@alejandro_ao]

hey there, that’s a super good question. i am making an entire video about prompt engineering with langchain very soon!

[@carlitosbal]

Hi Alejandro! It was an amazing tutorial, thank you for sharing. Do you know if its possible how to get a summarize a big amount of text or extract an idea of big amount of text? I saw that you ask simple questions, but how about a whole context interpretation?

[@agustn_pina]

hi Alejandro! great video! is there any way to determine the ideal chunk size?

[@gabrieljauregui654]

I am learning a lot of how LLMs work in the python ecosystem with your videos, also your solution-centered attitude helps me code less stressed when playing with such complex stuff haha, keep it up! 😃 💯

[@alejandro_ao]

i'm glad to hear that! keep it up and don't stop learning!

[@MindshiftMosaic]

Thanks a lot! As usual very clear and effective. If I can suggest a little but important extension is to show from which file and eventually the which page it found the answer. I think this could be a nice estensione of your previous video where you can add more than one document. Again, super compliments…My 2Cents

[@alejandro_ao]

@@MindshiftMosaic indeed, i had been planning to cover that in a video for a while! it will be making that soon :)

[@gabrieljauregui654]

@@MindshiftMosaic this is so cool! I was looking for a solution to that, I was even thinking in creating some rudimentary functions on my own but I didn't stop to think if there are some extensions already out there to solve that

[@MindshiftMosaic]

@@gabrieljauregui654 Based on my tests, without using additional external services, I would suggest one of these two ways to include additional information, such as the file name and page number, when transforming text into chunks and subsequently computing embeddings. Option 1: Extend the data structure representing each chunk or section. Instead of just representing a chunk as a plain string, you can define a class or data structure that encapsulates the content and the associated metadata. Option 2: You can store each chunk's file name, page number, and text in a database. ​​⁠​​⁠​​⁠@alejandro_ao do you have other options?

[@vinci_irl]

ALEJANDROOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOO

[@DSumForce]

Awesome video. Is there a way to narrow the token size of the documents accessed if you create your database out of something like 40 PDFs? I get errors saying that the context being passed to OpenAI is much greater than 4096 tokens...

[@kumarrohit3021]

Excellent piece of work…. This is what I was looking for…. Tons of thanks

[@alejandro_ao]

i’m glad it helped!

[@neginpirannanekaran1236]

This was awesome video. How can we change it to remember the previous conversations too?

[@TobiasOberrauch]

Looking forward to see qdrant

[@alejandro_ao]

let me know if you like it!

[@tristantisaja6807]

Wow that;s so cool application of vector database. Now with GPT 4 turbo, can it be combined with Quadrant or was it not necessary needed for chatting with documents since we have GPTs?

[@deepaksood5711]

Hey, my collection is not created showing Error: ResponseHandlingException: [Errno 99] Cannot assign requested address Can you help to resolve this error?

[@RedCloudServices]

Can you demonstrate this with PDFs using the new Streamlit Chat component? Also with all use cases, how do we remove a PDF from the vectorstore?

[@alejandro_ao]

hello there! sure thing, i’ve been using that component for a while and it’s super practical. i’ll post that soon

[@arslanabid2245]

But can you create a system where we upload a document and the chatbot takes help from the vector store, and explain according to that its knownledge base embeddings? is it possible ? if so, then please make a dedicated video on it

[@Jaypatel512]

I have a use case where we have list of "intents", and "their definition" explaining what that means. I want to make a QA bot that can answer intelligent questions about this dataset. What should be my strategy converting this data into a vector DB for proper effect ?

[@Tzonghannyang]

Thanks for the program. I am trying to combine read multiple pdf and save the embeddings in drant. If I can append the data.

[@mohitspg]

how can we add stream type response like openAI

[@alexpollock1710]

Thanks Alex... Any idea if you'll create a community or news letter?

[@BodrumDrone]

your video answers my question! What is the vector database? I want to create a location-based database for the city council district travel and food amenities. Thanks for sharing.

[@hardagerisanjay4042]

hello! i have been faced with a problem of not getting relationship data using vector databses. any chance of improving it for effecient retrieval ? or can we use different databses like sql, graph or two or more simultaneously for getting realtionship data.

[@AssassinUK]

Great video! Thanks, very well taught, your a good teacher. I'm using Flowise though to implement LangChain, a lot LOT faster plus it allows me to hit production a lot faster.

[@arkardtv]

The project works fine in my local tests. But when i upload to Azure i get the error "no module found: qdrant_client". Is there an extra step to make it work online?

[@NilsLuca]

Thank you very much for this great video! However, I have a question: How can I update the data? Suppose we want to add a new chapter to our beautiful story. Thank you very much for your help!

[@87nehal]

My data is overwriting the previous data is there any way to add more data to database without overwriting the previous one?

[@alejandro_ao]

sure thing! it seems like you are recreating the collection each time. try to just post the vectors using the add points endpoint

[@darshan7673]

Can we use any open source LLM instead of Open AI?

[@junaidlatif2881]

But how to upload our own database multiple pdf files?

[@nicholascameron_tf]

hi this is an excellent video, thank you. I have a question though, I am new at this so bear with me. In this example you first create a collection in the DB, then you split text into chunks and create the embeddings out of them (so the vector representation in OpenAi format) and then save the vectors into the collection. Then you use a Retrieval chain that uses OpenAI, which uses these vectors you have stored in the DB. My question are: Does openai charge me for using his LLM? Am I actually using the LLM behind the scenes? And the last question is, if the information i am asking is not in the text that has been stored in my DB, will OpenAI try to figure the response out? Do you know if the information stored in my DB is also shared with OpenAI? Thanks

[@kevinl.9657]

Yes, you are charged by OpenAI every time you embed. And also, you are charged every time you ask a question. So yes, you are actually using the LLM behind the scenes. I'm not sure about this but I think if you ask something that is not in the DB, then OpenAI will reply that it doesn't know the answer (maybe it depends on some factor but I saw some tutorial the OpenAI replied that it doesn't know the answer). I think the information stored in your DB is not shared with OpenAI, but I could be wrong.

[@oguzhanylmaz4586]

I am developing a chatbot that can work in an offline environment. I'm undecided about Vector db. Do you think Qdrant would be a good choice?

[@alejandro_ao]

hey there, totally. their stack is super solid and you can create a persistent local DB. an easier option (but with less features, i think) is Chroma

[@AIdevel]

Hi I lost my embeddings on qdrant tell how to save the embeddings on qdrant to persistant ??

[@alejandro_ao]

hey there, what do you mean you cave lost your embeddings on Qdrant? Were they stored in a Qdrant cloud instance or were you saving them in-memory? You can reply here or you can come to our discord server and ask a question in the forum! Here's the link: link.alejandro-ao.com/discord

[@AIdevel]

@@alejandro_ao no I created a collection after having initiated the cluster and store the text embeddings but during the process of working of working the session terminated I mistakenly rerun the cells including that creates the collection so it removed all the points and started again , my question is how can I combine multiple clusters in a single application If I have intermittent embeddings session la with different collections?

[@alejandro_ao]

@@AIdevel oh, i see! i would think that this is something that you have to deal with inside the code itself. you should be able to test wether or not a collection exists before sending a PUT request to Qdrant to create it. try to create a function like `collection_exists(collection_name)` that returns a boolean. and then only create the new collection if this function returns false

[@AIdevel]

@@alejandro_ao thank you so much 😊

[@yc6768]

Great video. I have a question: if i restart the python program, isn't it doing the embeddings again since the get_vector_store function needs to rerun? Is there a way to check whether the embedding exists already?

[@alejandro_ao]

yes, you would have to test if the text has already been processed before running the embeddings model. there are many ways to fo this. one is: you can add metadata to qdrant points to control which data has already been processed

[@DemoGPT]

Kudos on the excellent video! Your hard work is acknowledged. Could we expect a video about DemoGPT from you?

[@watchmen9773]

Thanks!

[@alejandro_ao]

Thank you!! You are the best

[@AIenthusiast-sn4gk]

Thank you so much for a very nice tutorial. Your clear explanation about each concepts helped to understand the concepts quickly. I have a question also. Say for example, You created embeddings from a text file and stored it in your vector space. But after some months, if you try to create embeddings for the same text file and try to store it in your same vector space. Will it throw any error? or Is there any way to validate this in order to avoid saving same data again and again?. It would be really helpful if you explain this. Looking forward to see more useful videos from you and thanks on behalf of everyone to your gesture.

[@red33devils]

I watched the videos of chatting with pdfs. When I test the models in HuggingFace without using OpenAi, I get incomplete and truncated answers. What could be the reason for this? I'm new to this field.

[@noasswa2010]

Context length?

[@red33devils]

@@noasswa2010 you mean chunk of length or my question of length

[@chrisalmighty]

@@red33devils I mean chunk length. Most models on Hugging Face 🤗 have short context so in a chat session, when it gets long, it can start to truncate the response due to length of the chunks sent and the response

[@red33devils]

@@chrisalmighty so, how can I solve this problem. Is it possible to implement "contiune editing" like chatGPT

[@muhammadzakiahmad8069]

Hey Alejandro loving your channel and videos so far, please make a series on MLOps if u can. Thankyou in Advance.

[@alejandro_ao]

hey there, thanks for the idea! i'll definitely do a video about deploying LLM applications 🔥

[@codelerner-tf8zi]

Great channel! This might be a silly question but if I build my own project based on your tutorial, can i do that? If yes, how should i credit your work, given that your github repos don't specify a license.

[@echo0204]

What about weaviate?

[@alejandro_ao]

Weaviate is really good too and they have a free sandbox in their cloud. I will probably make a video about them too 💪

[@echo0204]

@@alejandro_ao Thanks I am enjoying your tutorial! Awesome stuff. Do you know how we can get multiple answers (with source links) from several pdf embedding docs ? I am thinking about creating parallel prompt by creating several prompt variations in the background. But it doesn't seem to be effective. Wondering what's your thougt about it. Thanks!

[@timmymanamperi]

thank you heaps for another awesome video, this is very helpful video alternative to pinecone. thanks

[@DJPapzin]

How many indexes can i have?

[@dawidzywica2591]

Mb next video about Di id API and Generated video from photo? :)

[@Shsrif777]

Hi, you can Mp , prive ?

[@VicelsCollections]

You earned a subscirber

[@BABARKHAN-tc4dm]

kindly make video on how to deploy llma-2 on aws with endpoint

[@alejandro_ao]

i like this idea. expect this soon, i-m back at the studio lml

[@jeffg56]

Subscribed amazing videos!

[@alejandro_ao]

Welcome aboard!

[@levelupkareem]

could you do a tutorial on how to build an AI chatbot to use on a website. (Without Botpress and similar websites)

[@vladimirolezka3482]

Splendid video. Makes Langchain more appealing to use and learn.❤ Please how do I implement a different LLM(like Zephyr 7B) other than OpenAI? Thanks.

[@bendingreality-]

Great video, thanks for sharing! Is there a way to setup Qdrant quickly and without coding? I'd need to use it in combination with Flowise but even just following your instructions here makes my brain explode lol

[@Sp1Dski]

Great video! Keep up the good work! I am new to this AI programming and I am currently trying to combine 2 of your videos to make an app. This Video and also the video: "Langchain PDF App (GUI) | Create a ChatGPT For Your PDF in Python" I am trying to create an app that has an upload page, where I upload documents as I go and then the other page is where I chat with the AI about the documents I have uploaded. I don't want to manually input every new document I have, I just want to drag and drop the document on the upload page, hit "Upload" and then it should be in the database. However, I am struggling to accomplish this. Can you maybe help me out with a website I should take a look at so that I can figure this out? I would greatly appreciate it! Thank you!

[@AIenthusiast-sn4gk]

@Sp1Dski. Hi. I am also trying to upload files as well as chat with them. Did you achieve this already? It would be really helpful if you share some of your thoughts

[@ArunKumar-bp5lo]

openai embeddings not free right??

[@alejandro_ao]

right! you can always use open source embeddings models from huggingface

[@ergurkha3157]

Extraordinario.. gracias!

[@udaygupta5075]

Hi, can you make more such content with open source LLMs instead of ChatGPT

[@Guyassal.education]

thank you very much!!! 😀

[@alejandro_ao]

you’re very welcome :) keep dropping by for more

[@aayushiverma2290]

Hey!! Can u please suggest how can i integrate this with Microsoft teams Btw great video😉

[@AJ-ep4nj]

What happened no new video? Looks like you got something big and moved on! All the best

[@rafaelmartinsdecastro7641]

Good stuff

[@irfansaeedkhan7242]

its been 5 months why not new video, you got a job bro :p

[@alejandro_ao]

i was in a very time-consuming gig! now im back and will be doing this full time. no more big gigs got a good while 👌

[@watchbro3319]

But open ai api key is paid anyways

[@haiderkhalilpk]

useful

[@AssassinUK]

The free tier is not as sweet as it was a while back.

[@DIY_Foodie]

please make tutorial using of free llms api rather than openai 😓😓😓😓😓🙌🙌🙌🙌

[@S-Lomar]

😍😍😍🥰💕💕💕💕💕💞💞💞💞💞💞💞💞

[@faizrasool7990]

Please go back to IDE!!!

[@JakubSK]

Qdrant class doesn't exist.

[@ArunKumar-bp5lo]

# create the chain to answer questions qa_chain_instrucEmbed = RetrievalQA.from_chain_type(llm=OpenAI(temperature=0.2), chain_type="stuff", retriever=retriever, return_source_documents=True) here what's the alternative to openai i can't find any

[@paugonzalez5014]

Help, I have this error in the same code : vector_store = Qdrant( TypeError: Qdrant.__init__() got an unexpected keyword argument 'embeddings'

[@OmegaVestoLord]

did you ever fix this?

[@alexandrecgoes]

First of all, nice video! Helped me a lot! I have a problem here, I created a vectorstore using a big text of legal manual, around 250 pages, converted to txt. At the middle of it, there is a list of 10 cases that can occur in a specific situation. When I ask "Which cases can occur in this specific situation", I get only the first 3 cases. Increasing chunk_size and overlap to 3000 and 500, and LLM to ChatOpenAI to use model gpt-3.5-turbo-16k I could increase it to 7 cases. Increasing even more chunk_size/overlap got me worse (hallucinations) results. Any other ideas of how to solve this? Is there a way to see the return of the vectorstore query that is sent to the OpenAI to see if all the 11 cases are being sent? Btw, just a small correction in your notebook... You're using os.environ['QDRANT_COLLECTION'] when setting and os.getenv['QDRANT_COLLECTION_NAME'] when creating the vectorstore.

[@alexdominguess]

It is important to properly process the data before inserting it into the database. He have done is very simple example with a very small text. Instead of splitting the text by the number of char. you can split it by a special char like "#" So you go to every paragraph in your text and add "#" at the end to make sure sure that each chunk is not missing any part of its information.

[@MrAdityashah82]

@alejandro_ao great video! Could you possibly make a video on semantic search with langchain it gets very confusing.

[@huntingtonfreehan4579]

😂 Promo-SM

[@alejandro_ao]

nope, i wasn’t paid a cent for this