LlamaIndex
150,196 followers
- Report this post
QA over massive tabular data without hallucinations (Caltrain schedule edition šš°ļø)Even as LLMs get better, they still hallucinate over very complex tables and charts in documents - due to poor parsing.Case in point is the Caltrain schedule (or any big train schedule) - thereās a lot of time information in here! With LlamaParse, we were able to spatially layout the text in a semantically coherent manner, so that our GPT-4o-powered QA pipeline could correctly answer questionsš”In contrast with naive parsing (pypdf), a lot of tabular information gets lost, leading to LLM hallucinations.Check out our brand-new notebook here: https://lnkd.in/geCGEMbs
194
7 Comments
Hunter Zhao
Founder at Petal.org & GPT-trainer.com
1d
- Report this comment
Are you representing the table as markdown, csv, html, json, or some other format after extraction? There has to be some kind of limit to the tokens you can fit in a single chunk right? Since tables have such variety different formats, how do you ensure the data is stored according to the intended table structure?Our team at GPT-trainer studied this problem quite a bit. Hereās our approach: https://guide.gpt-trainer.com/working-with-tablesMicrosoft research also talked about LLM and tables: https://www.microsoft.com/en-us/research/blog/improving-llm-understanding-of-structured-data-and-exploring-advanced-prompting-methods/Then thereās the problem with multi-page tables too, and tables with merged cells, etc. that I assume are out of scope.
1Reaction
Titas Das
Applied ML Scientist | Software Engineer
1d
- Report this comment
How do we still keep the parsing quality high but rely less on a powerful LLM (because the cost can exponentially rise as the number of docs increase) ? Are there potential solutions from computer vision - such as segmentation, deep learning based object detection that can be integrated here to solve the problem of parsing not just tables but also plots and graphs while preserving the format?
1Reaction
Raj Kannan
AI Solutions Architect
1d
- Report this comment
THIS IS HOT! https://shorturl.at/oZzZY
1Reaction
Abdelfettah Latrache
software engineer at SYNDIKAT7
1d
- Report this comment
This is great!
1Reaction
Muhammad Habib
Global Investment Leader in Private Equity & Real Estate | FinTech Expert | $1B+ in Transactions
9h
- Report this comment
Muhammad Haseeb
1Reaction
Kabeer Singh Thockchom
AI & Data @ EY | AI Engineering + Product Management | Turning Ideas into Impact with AI: Creating Lasting Value from Ideation to Production | Azure; SaFe POPM
1d
- Report this comment
LlamaParse for the win!
1Reaction
To view or add a comment, sign in
More Relevant Posts
-
LlamaIndex
150,196 followers
- Report this post
Text-to-SQL - fully local edition šThe latest local LLMs are not only capable of RAG synthesis, but also querying structured databases. Diptiman Raichaudhuri has a great tutorial on how to build a fully local text-to-SQL setup, letting you query local databases without an internet connection. Stack:š¤ DuckDB as the databaseš¦Ollama + Mixtral-8x7B as the modelš¦ LlamaIndex as the text-to-sql orchestrationCheck out the full tutorial here: https://lnkd.in/gi_J7eNB
62
7 Comments
Like CommentTo view or add a comment, sign in
-
LlamaIndex
150,196 followers
- Report this post
Multi-Document Agentic RAG using LlamaIndex and MistralThis is a great article by Plaban Nayak on how to build a multi-document agent that can reason about multi-part questions over multiple documents.It does this by modeling each document as a set of tools (summarization and vector search). Since there can be many documents (so tools will overflow context), we can do tool retrieval first in order to pre-fetch the relevant tools that the agent will operate over. The end result is an advanced agent powered by Mistral AI function calling that goes many steps beyond what naive RAG systems can do. https://lnkd.in/gcuMfJJP
399
7 Comments
Like CommentTo view or add a comment, sign in
-
LlamaIndex
150,196 followers
- Report this post
Get 32x faster performance on your vector search at only a 4% cost in accuracy!Building production apps is all about tradeoffs. In vector search, your data is encoded as 32-bit vectors, which can use a lot of storage and compute to search. In this blog post from Jina AI, they show you how to get dramatically faster vector search by reducing them to binary digits, at a small cost in accuracy of retrieval.They also demonstrate how to get it working in LlamaIndex: it's as simple as adding a single `encoding_queries='binary'` parameter to your embedding call!https://lnkd.in/e4P82Dwk
136
4 Comments
Like CommentTo view or add a comment, sign in
-
LlamaIndex
150,196 followers
- Report this post
Structured Image Extraction with GPT-4o š¼ļøGPT-4o is state-of-the-art in integrating image/text understanding, and weāve created a full cookbook showing you how to use GPT-4o to extract out structured JSONs from images. It does this much better than GPT-4V. We feed it detailed papercards (created by Val Andrei Fajardo) of various research papers (see diagram below), and measure quantitative metrics like failure rates and quality of extracted outputs. GPT-4o was able to extract structured output from every image (0% failure rate), and synthesizes much higher quality answers/insights than 4V. Check out our full cookbook here - also by Val Andrei Fajardo!https://lnkd.in/gptzdcUq
406
20 Comments
Like CommentTo view or add a comment, sign in
-
LlamaIndex
150,196 followers
- Report this post
Announcing our first-ever meetup at our brand-new office in San Francisco! Come hang out with us and hear from us and our friends at Activeloop and Tryolabs about the latest developments in generative AI.https://lnkd.in/d3HXb3wF
96
2 Comments
Like CommentTo view or add a comment, sign in
-
LlamaIndex
150,196 followers
- Report this post
š„ Introducing GPT-4o + LlamaParse š„GPT-4o is the state-of-the-art model for multimodal understanding, meaning it also has state-of-the-art document parsing capabilities.LlamaParse is the platform for enabling LLM-powered parsing - it uses LLMs to extract documents from any file type in a performant, reliable fashion, offering state-of-the-art response quality for advanced document RAG.Weāre excited to offer GPT-4o as an explicit option in LlamaParse, which will use GPT-4o for extraction per page into markdown, instead of using our default parsers/models. Why:- GPT-4o is very good at parsing very complex documents into well-formatted markdown. Oftentimes it outperforms our default approaches.- This means that it can turn documents with very complex tables / charts into clean, indexable data for your RAG pipeline - higher response quality, lower hallucinations šTradeoffs / Caveats ā ļø:- Itās expensive šµ: Due to the cost of inference, using GPT-4o is currently $0.60 USD per page (while by default LlamaParse is $0.003 per page). This cost can spike quickly - beware!- You can specify your OpenAI key, in which case the marginal cost per page goes down to 0.3c per page.- This is a beta feature. Given the cost and latency, use this with caution! If you want to give this a shot, signup for an account and check out our UI: https://lnkd.in/gbkxQAQdNotebook: https://lnkd.in/grwUVr-G
652
23 Comments
Like CommentTo view or add a comment, sign in
-
LlamaIndex
150,196 followers
- Report this post
LlamaParse š¤ Quivr š§ Quivr (YC W24) (Stan Girard) is a popular open-source platform where you can create personalized AI assistants over your data.Weāre excited to partner with Quivr to introduce LlamaParse - parse any complex document (.pdf, .pptx, .md) through our advanced parsing capabilities, ensuring that you get clean data before storing in the agentās personalized memory. This ensures accurate retrieval and lower hallucinations during conversation.Check out the docs! https://lnkd.in/d7Ftjd5w
220
8 Comments
Like CommentTo view or add a comment, sign in
-
LlamaIndex
150,196 followers
- Report this post
GPT-4o support now available in create-llama! Get 90% of the way through building a chatbot over your data just by answering a few questions.
105
3 Comments
Like CommentTo view or add a comment, sign in
150,196 followers
View Profile
Follow