I built this tool to solve a problem that drove me crazy during my studies: course materials are often in formats that are terrible for quick review and reference. I was spending hours watching long video lectures before finals, trying to make sense of sparse PDF slides, or deciphering the ancient hieroglyphs in my friends' (or my own) notes.
The Lore Engine is a Python-based tool that processes content (from a single file or an entire folder) and generates comprehensive markdown files. You can feed it PDFs (slides, scanned notes), or transcripts with optional video context.
What makes it different is that you can control the output based on your needs. Through the CLI, you can choose:
- Content Type: (e.g., Slides, Textbook, Videos)
- Output Format: Detailed notes, practice problems, or a concise pre-exam revision sheet.
- Detail Level: From short-hand summaries to a deep-dive.
(Note: The practice problem and revision sheet modes are still experimental.)
The goal is to turn 10 hours of passive video watching into a ~2-hour focused read, while preserving all the necessary detail.
On the technical side:
It's a multimodal pipeline using the Gemini API (I chose it for the free tier and native multimodal support). The prompt system is modular, allowing for fine-grained control over the output style.
For video processing, it uses video_reader-rs (Rust FFmpeg bindings) instead of OpenCV for a significant performance boost and better memory management (it doesn't load entire videos into RAM).
To avoid redundant images (like 50 identical slides), it uses perceptual hashing (pHash) combined with a custom diversity scoring algorithm to select visually distinct frames.
The pipeline is multi-processed and uses round-robin API keys to scale throughput—which is essential when you're trying to process an entire course the night before an exam. It also has built-in rate limiting and exponential backoff for retries.
I've vouched for your comment. I don't see any reason why your comment would require vouching, so I would guess that some overzealous spam filter flagged it. I would recommend emailing hn@ycombinator.com to ask about the status of your account. If there's a problem that should take care of it. (You may have to go to your profile and add it to the `email` field first. This field isn't public but demonstrates your email is linked to your account.)
Not the warmest welcome to the community but congrats on shipping your project! It's an interesting problem space and I think some people could benefit a lot from such tools.
I really appreciate the vouch and the warm welcome! I was confused when the comment disappeared, thought I'd done something wrong. It was probably my inactive account. Will email hn@ycombinator.com to make sure everything's clear with it.
Thanks for the encouragement on the project! The problem space definitely resonated with me as a student, and I'm excited to see it helping others avoid the lecture-rewatching spiral. I made sure this was extendible enough that I can add more input formats (research papers, business documents) and output formats and tools in a few lines of code.
If, in the generated notes, you could link back to the original source (timestamp in video, slide number, etc.), I would use this all the time. I love LLMs but when extreme precision matters, like when studying for tests, I would love to be able to quickly reference the original source to verify the information myself.
Thanks! I actually had the same problem. The source linking is actually already there - just implemented a bit differently by directly inlining source images at relevant points:
For videos: Screenshots are timestamped in the filename as (frame_00-12-45.jpg)
For PDFs: Page numbers are preserved in the filename as (page_15.jpg)
The "verify as you study" use case is exactly what I built this for. I wanted the AI-enhanced explanations but still needed to double-check against source material during exam prep. The screenshot integration was designed to make that workflow natural instead of bouncing between 2 files.
That said, I could make this more explicit - maybe add the timestamps/page numbers as clickable links in the markdown itself. Would that be more useful than the current inline image approach? It will break if original file is moved.
Hey HN,
I built this tool to solve a problem that drove me crazy during my studies: course materials are often in formats that are terrible for quick review and reference. I was spending hours watching long video lectures before finals, trying to make sense of sparse PDF slides, or deciphering the ancient hieroglyphs in my friends' (or my own) notes.
The Lore Engine is a Python-based tool that processes content (from a single file or an entire folder) and generates comprehensive markdown files. You can feed it PDFs (slides, scanned notes), or transcripts with optional video context.
What makes it different is that you can control the output based on your needs. Through the CLI, you can choose: - Content Type: (e.g., Slides, Textbook, Videos) - Output Format: Detailed notes, practice problems, or a concise pre-exam revision sheet. - Detail Level: From short-hand summaries to a deep-dive.
(Note: The practice problem and revision sheet modes are still experimental.)
The goal is to turn 10 hours of passive video watching into a ~2-hour focused read, while preserving all the necessary detail.
On the technical side: It's a multimodal pipeline using the Gemini API (I chose it for the free tier and native multimodal support). The prompt system is modular, allowing for fine-grained control over the output style.
For video processing, it uses video_reader-rs (Rust FFmpeg bindings) instead of OpenCV for a significant performance boost and better memory management (it doesn't load entire videos into RAM).
To avoid redundant images (like 50 identical slides), it uses perceptual hashing (pHash) combined with a custom diversity scoring algorithm to select visually distinct frames.
The pipeline is multi-processed and uses round-robin API keys to scale throughput—which is essential when you're trying to process an entire course the night before an exam. It also has built-in rate limiting and exponential backoff for retries.
The project is open source and runs locally. Here is the link to the repository: https://github.com/Slydite/lore-engine
I'd love to get your feedback. Thanks for checking it out!
I've vouched for your comment. I don't see any reason why your comment would require vouching, so I would guess that some overzealous spam filter flagged it. I would recommend emailing hn@ycombinator.com to ask about the status of your account. If there's a problem that should take care of it. (You may have to go to your profile and add it to the `email` field first. This field isn't public but demonstrates your email is linked to your account.)
Not the warmest welcome to the community but congrats on shipping your project! It's an interesting problem space and I think some people could benefit a lot from such tools.
I really appreciate the vouch and the warm welcome! I was confused when the comment disappeared, thought I'd done something wrong. It was probably my inactive account. Will email hn@ycombinator.com to make sure everything's clear with it.
Thanks for the encouragement on the project! The problem space definitely resonated with me as a student, and I'm excited to see it helping others avoid the lecture-rewatching spiral. I made sure this was extendible enough that I can add more input formats (research papers, business documents) and output formats and tools in a few lines of code.
This is WILD. I love it. Congrats on shipping!
Thank you! Shipping for the first time was definitely nerve-wracking. Really appreciate the positive feedback!
Amazing idea and execution.
If, in the generated notes, you could link back to the original source (timestamp in video, slide number, etc.), I would use this all the time. I love LLMs but when extreme precision matters, like when studying for tests, I would love to be able to quickly reference the original source to verify the information myself.
Thanks! I actually had the same problem. The source linking is actually already there - just implemented a bit differently by directly inlining source images at relevant points:
For videos: Screenshots are timestamped in the filename as (frame_00-12-45.jpg)
For PDFs: Page numbers are preserved in the filename as (page_15.jpg)
The "verify as you study" use case is exactly what I built this for. I wanted the AI-enhanced explanations but still needed to double-check against source material during exam prep. The screenshot integration was designed to make that workflow natural instead of bouncing between 2 files.
That said, I could make this more explicit - maybe add the timestamps/page numbers as clickable links in the markdown itself. Would that be more useful than the current inline image approach? It will break if original file is moved.