José Manuel de la Chica’s Post

📈 LangGraph excels at creating custom processing chains, empowered by the LangChain Expression Language. Intriguingly, a common pattern emerges in complex LLM applications: the incorporation of cycles during execution. These cycles leverage the LLM's reasoning capabilities to determine the next step. This ability to reason within a loop is fundamental to unlocking the full potential of LLMs, essentially creating "agents." The power of this cyclic, "agentic behavior" is evident in Retrieval Augmented Generation (#RAG) applications. Traditionally, RAG applications retrieve documents and feed them to an #LLM for response generation. While effective, this approach falters when the initial retrieval fails to yield relevant results. A multi-agent approach can be significantly enhanced by leveraging #GraphRAG. This framework allows each agent to specialize by working with a domain-specific knowledge graph, leading to improved performance. FalkorDB is a cutting-edge graph database designed to power large language models (LLMs) with the fastest and most efficient knowledge infrastructure.

🤘🏼 Looking forward to my "Top Patterns for integrating Language Models into your own architectures" session at the #ittage next week. #GenAI #PracticalExperience #LanguageModels #EmbeddingModels #LLM #SLM #RAG #StructuredOutput #SemanticRouting #Observability

Monocle is a new reverse engineering tool from James Stevenson that combines local LLM with a bridge to Ghidra to implement a natural language search capability that will return candidate functions along with a match score and a high-level synopsis of what the function does: https://lnkd.in/e3rSes7Z (Github) The screen grab shows the output of searching for "authentication code" over pure-authd. A very neat foundation for further GenAI hackery.

🎉Exciting News🎉 Our paper, "Plaintext-Ciphertext Matrix Multiplication and FHE Bootstrapping: Fast and Fused," has been accepted at Crypto 2024 and will be presented this week! Our work delves into plaintext-cypthertext matrix multiplication, which is a key component of transformer-based large language models. Check out our full paper at: https://lnkd.in/gT_5fff7 #Crypto2024 #PrivacyPreservingML #FHE #LLM #MachineLearning

Plaintext-ciphertext matrix multiplication in less than 3.5 times the cost of plaintext-plaintext matrix multiplication! We also halve bootstrapping cost when several ciphertexts have to be bootstrapped in parallel. This is joint work with Youngjin Bae, Jung Hee Cheon, Guillaume Hanrot and Jai Hyun Park. Jai Hyun will be given the talk this Thursday at Crypto.

Do you really need all the fancy layers in LangChain to build a RAG system? In my experience, not always. LangChain introduces abstractions that can complicate things—making the logic less transparent and harder to improve in specific areas. In such a fast-evolving AI field, keeping it simple and clear often works better. That said, I’ve found LangChain’s text splitter and document loaders incredibly useful for document chunking without the abstraction downside above. But ultimately, to build a RAG system, all you need is: An LLM API/model A vector database with keyword and vector search Your code/functions/tools How do you approach building RAG systems? Share your thoughts and experience below, I’d love to hear your perspective.

Understanding why large language models (LLMs) make certain decisions can be challenging, but attributing importance to individual tokens within input prompts offers a powerful lens into their inner workings. In the context of time series, attribution methods like SHAP often struggle due to the abstract nature of single data points. However, when it comes to language, the interpretability of individual words provides clearer insights into model behavior. You can find the full paper on arXiv: https://lnkd.in/gtn-tPJd and github https://lnkd.in/gDrRtVuq 

Breaking Down Splitters in LangChain: Types and Their Functions Explained “Dissecting the splitters in LangChain is like unraveling a puzzle: each piece reveals a new pathway to seamless language understanding.” #splitters #typesofsplitters #langchain #langchaintool #langchainapps https://lnkd.in/dwe9AHXE

Recently I had the opportunity to work on LLM Project, I was amazed by the power on Lang chain and LLM Following is simple explanation of the steps in order to create ChatGPT in python using: * LangChain which is framework for developing applications powered By Language Models * huggingface_hub library allows you to interact with the Hugging Face platfrom which has thousands of models, datasets, and easily collaborate with ML workflows (it is free unlike OpenAI and it requires a HuggingFace Access Token) First: load the required documents you want to get information from using TextLoader function Second: Split each document to chunks with characterTextSplitter function Third: Convert the chunks to embedding (high-dimensional vectors represent each word or token) by calling HuggingFaceEmbeddings Forth: Store the embedding into vectorstore-DB Fifth: Call LLM model available in huggingFace(you can check in their official website (https://lnkd.in/dwTfsGBY) Sixth: Search the Similarity (by calling similarity search function) between user query and embedding stored in DB and use this similarity to generate an answer to the user question. #Edit As Example I load a document about Alcatraz Island and ask the model this question: How to reach Alcatraz Island? The mode answer was: You need to take a ferry from San Francisco Bay to the Island. and in different running its answer was: By a ferry Note: there is a limitation of Model's size, it should be less than 10 GB, also consider the model language when choosing one.

Omri Levy introduces an implementation to achieve sentence-window retrieval, which is a useful retrieval technique used in many RAG applications. Read the full article now! #LangChain #GenAI