Hey, thanks so much for the kind words and taking a look!

Search observation: What you're noticing with search is the currently intended behavior, but you raise an interesting point. I'm using tsvector for Postgres full-text search over title, abstract, etc. with prefix matching, specifically in English. The stemmer doesn't decompose Sanskrit so "brahmavihara" and "vihara" are separate tokens matching legitimately separate papers. Then it falls back to the embedding geometry to search by meaning instead of strings.

That said, there's enough Sanskrit going on in contemplative papers that a synonym/variant layer for Pali and Sanskrit vocabulary is a pretty good idea. There's a lot of German and Portugese in the corpus as well, so upgrading the search to account for all these languages is a great feature idea.

Tech question: There's actually zero LLM involvement at all regarding citations or references in this project. That all comes from OpenAlex API. LLMs only do relevance gating, study detail abstraction, and evidence synthesis summaries.

Google has a moat in the sense they have no API and don't license anything to anyone, so nobody builds on it and the open ecosystem of sources I'm using is the same as what AI products use. Gemini is trained on the same sources this library uses (and recently, this library itself) and the Google Scholar vault is in my opinion, pretty useless given the open ecosystem available.

Research question: the most mind-blowing results I've discovered are in an unexpected direction - it is wild how thin and low-confidence the evidence is for basically any claim regarding psychedelics, due primarily to functional unblinding and low sample size and diversity. The hype cycle is way beyond science can keep grounded.

Here's an example (disclaimer, it's my own published research!): A recent study reported an effect size for mebufotenin exceeding the theoretical maximum for antidepressants, implying methodological issues account for the results. https://consciousnesslibrary.org/research/52248-blinding-int...

Another "mind-blowing" observation is the sheer volume of consciousness work outside the psychedelic hype cycle. A full half of the corpus is on meditation or philosophy of mind, not related to psychedelics. And we still are so far from understanding the "hard problem.

Development question: I have plans to extend this indefinitely, and your factoid idea is interesting. The "study at a glance" might have most of what you're asking about already on each article, but taking that a step further, it could be cool to have a cross-corpus browsable feed of those. I'll also be integrating data from a project by Josie Kins (maker of https://effectindex.com) on psychedelic substance topics in particular for dose info and potentially trip reports, but that's a little orthogonal to what you're suggesting.

Let me know if you mean something more about factoids than the "study at a glance" and the potential to have those be aggregated in a browsable way and I'll see what I can think of.

I haven't bit off more than I can chew yet, my plan is to work on it until that's the case OR there's just nothing left to do in the scope of the project.

Thanks a lot for your comment!

Thank you! I couldn't agree more. Big questions to answer, and your framing is exactly why I went into this.

Thanks for taking a look! Yeah, these programs are very new and pretty uncommon (for now).

As for why I chose Rails - I just know Rails really well. I've been using it since 2012 or so, and it's basically muscle memory now. The most intensive parts of the app really would lend themselves to Python tooling and I did end up hand-building a lot of the data engineering stuff that there for sure exist plugins for in Python, but the trade off of using what I know well outweighed that by a lot. Plus the rest of it are perfectly CRUD-able server-rendered pages. Pretty Railsy aside from the data eng stuff.

I also started this before AI coding harnesses were available. I shelved it for a long time after the initial few rounds, but the work I did in the beginning was substantial enough to not just start over. If I had Claude Code with Fable when beginning, it might look a lot different. But I do like knowing everything about what I built and understanding every line.

On your AI ethics question... I do think it's possible, eventually. That's why I hedged a bunch on it being a long term goal. If open weight models gain steam and hardware acquisition becomes manageable, it would be a dream to run a self-hosted self-trained model in my house (which is powered by solar, so the renewable thing is real). It is especially realistic if some funding comes through for hardware.