Dominik Weckmüller

GIS & Data Science | Consultant @ & @EU_ScienceHub
PhD Student @tudresden_de
fosstodon & bluesky @DomeGIS

[Dominik Weckmüller @DomeGIS]

Did you know that OSM data includes "building levels"?
Following my tutorial you can create a 3D-map with #QGIS and #Aerialod and even emulate roofs!

https://t.co/LItd7jAkhk

#30DayMapChallenge day 8: OpenStreetMap.
#gischat

[Dominik Weckmüller @DomeGIS]

🌍#30DayMapChallenge
Day 29. Data: Overture – Multilingual Geospatial Semantic Search 🔍
Ever wanted to query places around the world in your local language?
E.g.
- "예술적인 장소" in Mongolia
- "Футбол" in Uganda
App: https://t.co/SGPaFbdW3H
#gischat

[Dominik Weckmüller @DomeGIS]

Query Foursquare's 100M OS places in a one-liner with DuckDB without downloading! https://t.co/MGj7aop3Po

[Dominik Weckmüller @DomeGIS]

Hacking the FlatGeobuf geospatial file format for building a single static file vector database. https://t.co/BdDnvowm5y

[Dominik Weckmüller @DomeGIS]

Calculate semantic similarity in your browser based on Excel or CSV tables with transformers.js & Minishlab's Potion/model2vec models! Semantic Similarity Table is highly performant and private, all data remains in your browser: https://t.co/KTimthulrb

[Dominik Weckmüller @DomeGIS]

SemanticFinder now supports WebGPU thanks to @Xenova's efforts!
Expect massive performance gains. Inferenced a whole book with 46k chunks in <5min!
https://t.co/IysEEpPzlN
If your device doesn't support #WebGPU use the classic Wasm-based version https://t.co/IyCl9CycS2

[will brown @willccbb]

wait, you still have to give it the system prompt???? lol

are the weights even different from normal Llama3-70B?

the published evals are completely indefensible then, whole thing smells like a grift

[Dominik Weckmüller @DomeGIS]

@varunneal @arpitingle @spikedoanz Absolutely! That's the main idea of the index collection here for #SemanticFinder: https://t.co/uRyNtRjjw8
You can create an index and keep it private or share with others. Theoretically one could also encode the entire index in the URL itself.

[Dominik Weckmüller @DomeGIS]

@geolytica Did you consider joining/merging the 52Mio Overturemaps places to poidata?

[Dominik Weckmüller @DomeGIS]

Super grateful for the opportunity to talk on @MapScaping about my research: Semantic Search for Geospatial - Let me know what you think! :)
Mapscaping: https://t.co/3n1dxwHaOq
Spotify: https://t.co/PKMHcgVyKy
My homepage: https://t.co/DINN3t6uPu
Demo: https://t.co/Kfa3278z2Y

[Dominik Weckmüller @DomeGIS]

@willccbb @levelsio Just created https://t.co/aqDuNr76om as a simple UI (back- & frontend) for this purpose. Add as many prompts as you like and inference in parallel. You could add a RAG-pipeline of your choice to improve the results.

[Dominik Weckmüller @DomeGIS]

Blown away by @willccbb's https://t.co/U1AO4s4WUM. Did you know you can get 10x the throughput if you inference in parallel on your Mac? Results for mlx-community/Meta-Llama-3-8B-Instruct-4bit. #mlx

[Dominik Weckmüller @DomeGIS]

Daily notifications about trending feature-extraction models compatible with transformers.js for semantic search! Sortable tables & scripts all open source built on GitHub Actions & https://t.co/AEZxJHvWyx.
https://t.co/T6L0B9eBsJ
145 models usable in https://t.co/FL6W0qGQD3!

[Dominik Weckmüller @DomeGIS]

The dots are colored by relevance to your search query, here "food" (dark blue dot). More relevant dots are darker, less relevant ones lighter.

[Dominik Weckmüller @DomeGIS]

The whole bible in 62.000 embeddings (or 31.000 verses) searchable and explorable in your browser with #SemanticFinder leveraging Barnes-Hut t-SNE. https://t.co/Q5870T5P4U. It takes a while to process (t-SNE is comp. intense) but it's smooth to explore on consumer-grade hardware.

[Dominik Weckmüller @DomeGIS]

You can quickly identify important characters like "Balak the king of Moab" with plenty of verses mentioning him.

The long chain of dots is Noah's story with his children.

[Dominik Weckmüller @DomeGIS]

I am using a simple regex to split the verses \{([^}]+)\} by verse number e.g. {6:1} so we get 62k embeddings instead of only 31k. The cluster you can see on top contains all these numbers. Interesting to see how the embedding model & t-SNE cluster numbers.

[Jina AI @JinaAI_]

WE R OPEN SAUCE!

https://t.co/IJYqBVy9sz

[Dominik Weckmüller @DomeGIS]

@0xa5ad @ollama @MistralAI Feel the same! With a local copy of SemanticFinder and an index of your choice you even have offline RAG capabilities in your browser - Ollama is integrated: https://t.co/oGC8j9sYdq

[Dominik Weckmüller @DomeGIS]

Interactively visualizing 169 #SDG targets with @JinaAI_ /jina-embeddings-v2-base-en, tSNE and #deckgl reveals interesting overlaps & synergies!
Discover here: https://t.co/byKxeudJxJ