Been working on this AI search for a while now, pretty much as a one-man army. Had to handle everything-setting up K8S, buying GPUs, figuring out storage, ensuring reliability with replicas, tweaking the network... you name it. Even built a full-text search from scratch, added OCR for PDFs, and managed chunking, embedding, storing, assessing... It’s like a billion tiny steps that finally led to Space Frontiers.
Oh, and why’d I call it Space Frontiers? Ever since I was a kid, I’ve been obsessed with the idea of creating something like the Borg-you know, a super-intelligence linked to all of humanity’s digital knowledge. I dreamed it’d guide colonial starships to other worlds, help create new life forms, and push consciousness out to the freaking edges of space. Yeah, that dream’s still a long way off, but I’m grinding to make it happen.
Right now, Space Frontiers search runs on just two servers, so it’s kinda slow. But for academic searches-things like Standards, PubMed, Reddit, Telegram, patents, books, and so on-it’s just a query away. The quality of answers seems better than tools like Perplexity in Academic mode, Elicit, or Consensus. It’s not perfect, and sometimes it flops, but other times the results are so cool that it’s become a daily go-to alongside other search engines. Give it a try; hopefully, it’ll impress.
It’s a regular RAG setup with a large database, query reformulation, and expansion. The databases are split into BM25 and vector parts - the first is powered by my own search engine, Summa, and the vector part runs on AlloyDB. Embeds are handled by Jina v3, while chunking and processing come from a ton of our boilerplate code. Things get reranked by a reranker, then chunks are combined, the document set is expanded and reranked again, and finally sent to LLMs with tuned prompts. That’s the gist of it. Oh, and Qwen3 is the king, by the way.
Starting to see how AI paired with a fleet of knowledge sources should be organized, and there’s a ton left to tackle. That’s why there’s a search for a co-founder, pre-seed funding, and anything else needed to turn this dream into reality. If interested, let’s talk.
pasha_sf•8mo ago
Oh, and why’d I call it Space Frontiers? Ever since I was a kid, I’ve been obsessed with the idea of creating something like the Borg-you know, a super-intelligence linked to all of humanity’s digital knowledge. I dreamed it’d guide colonial starships to other worlds, help create new life forms, and push consciousness out to the freaking edges of space. Yeah, that dream’s still a long way off, but I’m grinding to make it happen.
Right now, Space Frontiers search runs on just two servers, so it’s kinda slow. But for academic searches-things like Standards, PubMed, Reddit, Telegram, patents, books, and so on-it’s just a query away. The quality of answers seems better than tools like Perplexity in Academic mode, Elicit, or Consensus. It’s not perfect, and sometimes it flops, but other times the results are so cool that it’s become a daily go-to alongside other search engines. Give it a try; hopefully, it’ll impress.
It’s a regular RAG setup with a large database, query reformulation, and expansion. The databases are split into BM25 and vector parts - the first is powered by my own search engine, Summa, and the vector part runs on AlloyDB. Embeds are handled by Jina v3, while chunking and processing come from a ton of our boilerplate code. Things get reranked by a reranker, then chunks are combined, the document set is expanded and reranked again, and finally sent to LLMs with tuned prompts. That’s the gist of it. Oh, and Qwen3 is the king, by the way.
Starting to see how AI paired with a fleet of knowledge sources should be organized, and there’s a ton left to tackle. That’s why there’s a search for a co-founder, pre-seed funding, and anything else needed to turn this dream into reality. If interested, let’s talk.