Morphik – Open-source RAG that understands PDF images, runs locally
Hey HN, we’re Adi and Arnav. A few months ago, we hit a wall trying to get LLMs to answer questions over research papers and instruction manuals. Everything worked fine, until the answer lived inside an image or diagram embedded in the PDF. Even GPT‑4o flubbed it (we recently tried O3 with the same, and surprisingly it flubbed it too). Naive RAG pipelines just pulled in some text chunks and ignored the rest. We took an invention disclosure PDF (https://drive.google.com/file/d/1ySzQgbNZkC5dPLtE3pnnVL2rW_9...) containing an IRR‑vs‑frequency graph and asked GPT “From…
In plain words
Morphik is an open-source retrieval-augmented generation system designed to extract and answer questions from PDF documents containing images, diagrams, and embedded visual content. Built for researchers and professionals working with technical manuals and research papers, it runs locally and addresses the limitation of standard RAG pipelines that ignore images within PDFs. Unlike conventional approaches that struggle with visual information, Morphik processes both text and images to provide accurate answers to questions about document content.
written from the facts on this page · September 2026
From the sources
In the maker’s words, at launch
Hey HN, we’re Adi and Arnav. A few months ago, we hit a wall trying to get LLMs to answer questions over research papers and instruction manuals. Everything worked fine, until the answer lived inside an image or diagram embedded in the PDF. Even GPT‑4o flubbed it (we recently tried O3 with the same, and surprisingly it flubbed it too). Naive RAG pipelines just pulled in some text chunks and ignored the rest. We took an invention disclosure PDF (https://drive.google.com/file/d/1ySzQgbNZkC5dPLtE3pnnVL2rW_9...) containing an IRR‑vs‑frequency graph and asked GPT “From the graph, at what frequency is the IRR maximized?”. We originally tried this on gpt-4o, but while writing this used the new natively multimodal model o4‑mini‑high. After a 30‑second thinking pause, it asked for clarifications, then churned out buggy code, pulled data from the wrong page, and still couldn’t answer the question. We wrote up the full story with screenshots here: https://docs.morphik.ai/blogs/gpt-vs-morphik-multimodal. We got frustrated enough to try fixing it ourselves. We built Morphik to do multimodal retrieval over documents like PDFs, where images and diagrams matter as much as the text. To do this, we use Colpali-style embeddings, which treat each document page as an image and generate multi-vector representations. These embeddings capture layout, typography, and visual context, allowing retrieval to get a whole table or schematic, not just nearby tokens. Along with vector search, this could now retrieve exact pages with relevant diagrams and pass them as images to the LLM to get relevant answers. It’s able to answer the question with an 8B llama 3.1 vision running locally! Early pharma testers hit our system with queries like "Which EGFR inhibitors at 50 mg showed ≥ 30% tumor reduction?" We correctly returned the right tables and plots, but still hit a bottleneck, we weren’t able to join the dots across multiple reports. So we built a knowledge graph: we tag entities in both text and images, normalize synonyms (Erlotinib → EGFR inhibitor), infer relations (e.g. administered_at, yields_reduction), and stitch everything into a graph. Now a single query could traverse that graph across documents and surface a coherent, cross‑document answer along with the correct pages as images. To illustrate that, and just for fun, we built a graph of 100 Paul Graham’s essays here: https://pggraph.streamlit.app/ You can search for various nodes, (eg. startup, sam altman, paul graham and see corresponding connections). In our system, we create graphs and store the relevant text chunks along with the entities, so on querying, we can extract the relevant entity, do a search on the graph and pull in the text chunks of all connected nodes, improving cross document queries. For longer or multi-turn queries, we added persistent KV caching, which stores intermediate key-value states from transformer attention layers. Instead of recomputing attention from scratch every time, we reuse prior layers, speeding up repeated queries and letting us handle much longer context windows. We’re open‑source under the MIT Expat license: https://github.com/morphik-org/morphik-core Would love to hear your RAG horror stories, what worked, what didn’t and any feedback on Morphik. We’re here for it.
More ai this month
the category →
I trained a 125M-parameter transformer to autocomplete piano performances in real time (~108 notes/sec on an iPhone 15). The idea is basically GitHub Copilot or Tabnine, except instead of prompting it with code, you prompt it by playing a few notes on a MIDI piano. The model then continues what you played, entirely on-device. The app is free if anyone wants to try it. Happy to answer questions about the model, training, Core ML, or the many things that didn't work.
AI · 17d ago · simedw.com
Astute▲585Automate your B2B brand going viral, with new media creators
AI · 18d ago · company-app.joinastute.com
Hey HN, Henry from Cactus here! We previously released Cactus Needle, a 14MB agentic LLM for tool call, device use, and structured extraction for phones, wearables, smart homes, small robots and microcontrollers. We got really great feedback here, and have now incorporated the suggestions to release Needle 2. The whole model is a single 14MB binary that runs a full session in 28MB of RAM; 45m parameters at 2bit compression. Needle hits 500 tokens/sec decode speed on a Raspberry Pi 5, sits between 400-1,500 tokens/sec on VR devices like Meta Quest 3S and Apple Vision Pro, and ranges…
AI · 27d ago · cactuscompute.com
Launched alongside, April 2025
the whole month →- IB
Hi everyone, I built PyXL — a hardware processor that executes a custom assembly generated from Python programs, without using a traditional interpreter or virtual machine. It compiles Python -> CPython Bytecode -> Instruction set designed for direct hardware execution. I’m sharing an early benchmark: a GPIO test where PyXL achieves a 480ns round-trip toggle — compared to 14-25 micro seconds on a MicroPython Pyboard - even though PyXL runs at a lower clock (100MHz vs. 168MHz). The design is stack-based, fully pipelined, and preserves Python's dynamic typing without static type restrictions.…
Dev tools · 2025 · runpyxl.com
- UC
Life & fun · 2025 · filiph.github.io
- IB
https://the-pocket.github.io/Tutorial-Codebase-Knowledge/
AI · 2025 · github.com
