保留正文与补采附件。源页面当前未重新核验；具体限制见交付清单。

Revisit consent button We value your privacy We use cookies to enhance your browsing experience, serve personalised ads or content, and analyse our traffic. By clicking "Accept All", you consent to our use of cookies. Customise Reject All Accept All Customise Consent Preferences We use cookies to help you navigate efficiently and perform certain functions. You will find detailed information about all cookies under each consent category below. The cookies that are categorised as "Necessary" are stored on your browser as they are essential for enabling the basic functionalities of the site. ... Show more NecessaryAlways Active Necessary cookies are required to enable the basic features of this site, such as providing secure log-in or adjusting your consent preferences. These cookies do not store any personally identifiable data.

Functional Functional cookies help perform certain functionalities like sharing the content of the website on social media platforms, collecting feedback, and other third-party features.

Analytics Analytical cookies are used to understand how visitors interact with the website. These cookies help provide information on metrics such as the number of visitors, bounce rate, traffic source, etc.

Performance Performance cookies are used to understand and analyse the key performance indexes of the website which helps in delivering a better user experience for the visitors.

Advertisement Advertisement cookies are used to provide visitors with customised advertisements based on the pages you visited previously and to analyse the effectiveness of the ad campaigns.

Uncategorised Other uncategorised cookies are those that are being analysed and have not been classified into a category as yet. No cookies to display. Reject All Save My Preferences Accept All Toggle SearchToggle Mobile Navigation Publish AI, ML & data-science insights to a global community of data professionals. Sign inSubmit an Article LatestEditor's PicksNewsletterShipAI

Write For TDS LinkedInXToggle Search Large Language Models

Bringing Vision-Language Intelligence to RAG with ColPali

Unlocking the value of non-textual contents in your knowledge base Julian Yip October 29, 20257 min read Source: Image by the author. If you’ve ever tried building a RAG (Retrieval-Augmented Generation) application, you’re likely familiar with the challenges posed by tables and images. This article explores how to tackle these formats using Vision Language Models, specifically with the ColPali model. But first, what exactly is RAG — and why do tables and images make it so difficult?

RAG and parsing

Imagine you’re faced with a question like:

What’s the our company’s policy for handling refund? A foundational LLM (Large Language Model) probably won’t be able to answer this, as such information is company-specific and typically not included in the model’s training data. Learn this step by step with the interactive AI Engineer roadmap. That’s why a common approach is to connect the LLM to a knowledge base — such as a SharePoint folder containing various internal documents. This allows the model to retrieve and incorporate relevant context, enabling it to answer questions that require specialized knowledge. This technique is known as Retrieval-Augmented Generation (RAG), and it often involves working with documents like PDFs. However, extracting the right information from a large and diverse knowledge base requires extensive document preprocessing. Common steps include:

  1. Parsing : Parsing documents into texts and images, often assisted with Optical Character Recognition (OCR) tools like Tesseract. Tables are most often converted into texts
  2. Structure Preservation : Maintain the structure of the document, including headings, paragraphs, by converting the extracted text into a format that retains context, such as Markdown
  3. Chunking : Splitting or merging text passages, so that the contexts can be fed into the context window without causing the passages come across as disjointed
  4. Enriching : Provide extra metadata e.g. extract keyword or provide summary to the chunks to ease discovery. Optionally, to also caption images with descriptive texts via multimodal LLM to make images searchable
  5. Embedding : Embed the texts (and potentially the images too with multimodal embedding), and store them into a vector DB

As you can imagine, the process is highly complicated, involves a lot of experimentation, and is very brittle. Worse yet, even if we tried to do it as best as we could, this parsing might not actually work after all.

Why parsing often falls short

Tables and image often exist in PDFs. The below image shows how they are typically parsed for LLM’s consumption: Source: Image by the author.

However, there are two inherent issues with such traditional approach. #1. Complex tables cannot be simply be interpreted as texts
Taking this table as an example, we as human would interpret that a temperature change of **> 2˚C to 2.5˚C’**s implication on Health is A rise of 2.3˚C by 2080 puts up to 270 million at risk from malaria Source: The Impacts and Costs of Climate Change However, if we turn this table into a text, it would look like this: Temperature change Within EC target <(2˚C) >2˚C to 2.5˚C >3C Health Globally it is estimated that A rise of 2.3oC by 2080 puts A rise of 3.3oC by 2080 an average temperature rise up to 270 million at risk from would put up to 330... The result is a jumbled block of text with no discernible meaning. Even for a human reader, it’s impossible to extract any meaningful insight from it. When this kind of text is fed into a Large Language Model (LLM), it also fails to produce an accurate interpretation. #2. Disassociation between texts and images
The description of the image is often included in texts and they are inseparable from one another. Taking the below as an example, we know the chart represents the “Modelled Costs of Climate Change with Different Pure Rate of Time Preference and declining discount rate schemes (no equity weighting)” Source: The Impacts and Costs of Climate Change However, as this is parsed, the image description (parsed text) will be disassociated with the image (parsed chart). So we can expect, during RAG, the image would not be retrieved as input when we raise a question like “what’s the cost of climate change?” So, even if we attempt to engineer solutions that preserve as much information as possible during parsing, they often fall short when faced with real-world scenarios. Given how critical parsing is in RAG applications, does this mean RAG agents are destined to fail when working with complex documents? Absolutely not. With ColPali, we have a more refined and effective approach to handling them.

What’s ColPali?

The core premise of ColPali is simple: Human read PDF as pages, not “chunks”, so it makes sense to treat PDF as such: Instead of going through the messy process of parsing, we just turn the PDF pages into images, and use that as context for LLM to provide an answer. Now, the idea of embedding images using multimodal models isn’t new — it’s a common technique. So what makes ColPali stand out? The key lies in its inspiration from ColBERT, a model that embeds inputs into multi-vectors, enabling more precise and efficient search. Before diving into ColPali’s capabilities, let me briefly digress to explain what ColBERT is all about.

ColBERT: Granular, context-aware embedding for texts

ColBERT is a text embedding and reranking technique that leverage on multi-vectors to enhance search accuracy for texts. Let’s consider this case: we have this question: is Paul vegan?, we need to identify which text chuck contains the relevant information. Highlighted in yellow are texts which contain information about Paul Ideally, we should identify Text Chunk A as the most relevant one. But if we use a single-vector embedding model (text-ada-002), it will return Text Chunk B instead. The reason lies in how single-vector bi-encoders — like text-ada-002 — operate. They attempt to compress an entire sentence into a single vector, without encoding individual words in a context-aware manner. In contrast, ColBERT embeds each word with contextual awareness , resulting in a richer, multi-vector representation that captures more nuanced information. Numbers in the vectors are illustrative and do not represents the actual values

ColPali: ColBERT’s brother for handling document-like images

ColPali follows a similar philosophy but applies it to document-like images. Just as ColBERT breaks down text and embeds each word individually, ColPali divides an image into patches and generates embeddings for each patch. This approach preserves more of the image’s contextual detail, enabling more accurate and meaningful interpretation. Apart from higher retrieval accuracy, benefits of ColPali includes:

  1. Explainability : ColPali enables word-level comparison between the query and individual image patches of a document. This allows us to clearly understand and justify why a particular document is deemed more relevant.
  2. Reduced Development Effort & Greater Robustness: By eliminating the need for complex preprocessing pipelines — such as chunking, OCR, and layout parsing — ColPali significantly reduces development time and minimizes potential points of failure.
  3. Performance Gains : Embedding and retrieval processes are faster, resulting in improved overall system responsiveness.

Now you know what ColPali is, let’s dive into the code and see if ColPali can resolve the challenges we mentioned earlier!

Illustration

My code can be found in my Github. A few words about it:

Results

The code is pretty straightforward, so I will jump into conclusion instead: How well can ColPali deal with the problems we have seen in the “Why parsing often falls short?” section? #1. Complex tables cannot be simply be interpreted as texts Question : What’s the health impact if temperature rises above pre-industrial level by 3 degrees? Answer : If the temperature rises above pre-industrial levels by 3 degrees Celsius, the health impact would be significant. Specifically, it is estimated that by the 2080s, up to 330 million people would be at risk from malaria. This is a substantial increase compared to the 1920s, where the number at risk was between 50 to 100 million. This is correctly referred from this table in the PDF! Source: Page 3 of The Impacts and Costs of Climate Change #2. Disassociation between texts and images Question : What’s the modelled cost of climate change, according to the Greenbook declining model? Answer : According to the Greenbook declining model, the modelled cost of climate change is 7.2 Euro/tCO2 The correct answer should be 7.4 Euro/tCO2, but we can see it is close to correct! Source: Page 46 of The Impacts and Costs of Climate Change

Conclusion

Traditional RAG pipelines struggle with non-textual content. ColPali treats each PDF page as an image, allowing it to process visual layouts, tables, charts, and embedded graphics — formats that standard text parsers often distort or ignore. ColPali brings vision-language intelligence to RAG, making it far more capable of handling the messy, multimodal reality of enterprise documents. *Unless otherwise noted, all images are by the author

Register Now to Continue Reading

Unlock Towards Data Science with a Free Account

All articles on Towards Data Science are available to registered users. Create a free account today to access this article and thousands more. Register for freeSign in to your account

Written By

Julian Yip

See all from Julian Yip Editor's PicksArtificial IntelligenceGenerative AiMultimodalityRag Share This Article

Towards Data Science is a community publication. Submit your insights to reach our global audience and earn through the TDS Author Payment Program. Write for TDS

Related Articles

Detecting Insecure Code with LLMs

Large Language Models Prompt Experiments for Python Vulnerability Detection Melanie Hart Buehler March 21, 202411 min read

Running Llama 2 on CPU Inference Locally for Document Q&A

Large Language Models Clearly explained guide for running quantized open-source LLM applications on CPUs using LLama 2, C Transformers, GGML, and LangChain Kenneth Leung July 18, 20234 min read

Benchmarking time series datasets with style

How to benchmark/spot differences between time series datasets? M Adel October 20, 20213 min read

Topic Model Labelling with LLMs

Large Language Models Python tutorial for reproducible labeling of cutting-edge topic models with GPT4-o-mini. Petr Koráb July 14, 20256 min read

Highlights on Large Language Models at KDD 2023

Deep Learning Were you not able to attend KDD? Learn from my summary about the hottest topic at the conference: LLMs Gabriel Moreira September 11, 20238 min read

Heart Disease Prediction

Artificial Intelligence Cleveland Heart Disease(UCI Repository) dataset - classification with various models. Shubhankar Rawat August 10, 20198 min read

I Invented a Way to Speak to an AI, Keeping Your Privacy

Artificial Intelligence The tech is called "Silent Voice." Rafe Brena, Ph.D. June 27, 20248 min read

How to Enrich LLM Context to Significantly Enhance Capabilities

Large Language Models Learn how to empower your LLMs by leveraging additional metadata Eivind Kjosbakken September 16, 20257 min read

Text Watermarking in Python: Catch Whoever Copies Your Writing

Large Language Models AI companies quietly watermark billions of words a day. Here’s how to apply the same three families of techniques to your own writing—and what real experiments reveal about which watermarks survive copy-paste, editing, and paraphrasing. Chien Vu Minh September 6, 202626 min read

Your home for data science and AI. The world's leading publication for data science, data analytics, data engineering, machine learning, and artificial intelligence professionals. Subscribe to Our NewsletterWrite For TDS•About•Advertise•Privacy Policy•Terms of Use © Insight Media Group, LLC 2026•The New Stack•roadmap.sh•Creator Network

Recommended Articles

Close