MarkItDown, turbovec and LiteRT-LM
- Microsoft’s MarkItDown, Ryan Codrai’s turbovec and Google’s LiteRT-LM are three open-source tools tackling separate choke points in AI app plumbing. - MarkItDown turns files into Markdown, turbovec says 10 million vectors shrink from 31 gigabytes to 4, and LiteRT-LM runs models on-device. - Together they map ingestion, retrieval and edge inference work now moving into production stacks. (github.com 1) (github.com 2) (ai.google.dev)
Before a chatbot can answer a question, software has to turn files into text, store that text for search, and run a model somewhere. MarkItDown, turbovec and LiteRT-LM each target one of those steps. (github.com 1) (github.com 2) (ai.google.dev) MarkItDown is Microsoft’s Python tool for converting documents into Markdown, a plain-text format that keeps headings, tables and links. Its GitHub README says it is built for large language model and text-analysis pipelines rather than pixel-perfect document rendering. (github.com) (pypi.org) That matters because retrieval systems usually work on text chunks, not on raw PDF layouts or PowerPoint slides. MarkItDown’s latest GitHub release, version 0.1.5 on February 20, 2026, included PDF table extraction and list-parsing fixes aimed at cleaner downstream text. (github.com) (pypi.org) turbovec handles the next step: vector search, where text is turned into long lists of numbers and compared for similarity. The project says a 10 million document corpus that takes 31 gigabytes as float32 vectors can fit in 4 gigabytes with its index. (github.com) (pypi.org) Its pitch is speed without a training step. turbovec says it is built in Rust with Python bindings on Google Research’s TurboQuant method, and that new vectors can be added without codebook retraining or index rebuilds as the corpus grows. (github.com) (pypi.org) LiteRT-LM covers the last piece: running a language model directly on a device instead of sending every prompt to a cloud server. Google describes it as a production-ready, open-source framework for Android, iOS, web, desktop and Internet of Things hardware such as Raspberry Pi. (ai.google.dev) (github.com) Google says LiteRT-LM can use graphics processing units and neural processing units, the phone and laptop chips built for machine-learning math. The project’s GitHub repository says it is already used in Google products including Chrome, Chromebook Plus and Pixel Watch. (ai.google.dev) (github.com) Put together, the three projects sketch a common AI backend pattern in 2026: ingest messy files, compress and search embeddings locally, and keep some inference on the edge. That setup can reduce bandwidth, cut latency and keep sensitive documents off remote servers. (github.com 1) (github.com 2) (ai.google.dev) They also come from three different corners of the stack: Microsoft for document ETL, an independent Rust project for vector indexing, and Google for device runtime. The split shows how AI infrastructure is being assembled from specialized open-source parts rather than one monolithic platform. (github.com 1) (github.com 2) (github.com 3)
