Parabeagle Document Indexing
An AI replacement for Disinfodrome
I began working with Open Semantic Search in 2019 and the resulting service, Disinfodrome, was turned down at the end of January this year. This service had about two dozen users at its peak and they were accessing various facets of a collection of 844,000+ documents. Use fell to almost nothing after the 2024 election and it wasn’t worth maintaining something that only I used.
Fast forward to August of 2025 and Parabeagle has been named and made public. Let me show you around this document indexing Model Context Protocol server.
Attention Conservation Notice:
In this episode of Old Man Yells At Cloud, I will complain about Claude’s instability, while exploring fairly esoteric document indexing knowledge. This is just advisory, unless you’re technical enough to install a Python companion to Claude Desktop.
Parabeagle:
There are an enormous number of MCP servers available, over 1,100 on this official list, and another fifty are queued up waiting to be added. Once I get a chance to test Parabeagle on a Windows machine, it will join the herd. I even registered parabeagle.dog for the system, but I haven’t worked out the redirects yet.
The software is a fork of Chroma’s MCP server, but I’ve added features meant for those who work on court cases. These features include:
Multiple unconnected folders, so cases are kept separate, you pick which workspace you want to use from within Claude Desktop.
Document format aware chunking, a fancy way of saying the paragraphs stay together, instead of files getting chopped up every thousand characters.
Replaced 384 dimension embed with 768 dimensions, takes 50% longer, but much more accurate on finding stuff.
Loading documents via an MCP client like Claude Desktop is absolutely miserable, so it’s disabled in Parabeagle, and instead you have to use the speedy command line tools to do that.
Attribution of search results is a must for court cases, so Parabeagle includes a function that will show the file names and their paths when you search.
The big thing left to do is to provide a smooth way to export a collection from one system and import it to another. When the paralegal is done doing the import and review, counsel on the case will want direct access.
Claude Wobbles:
During the time I first started working on this software Claude was having a LOT of performance problems, often being completely unavailable. Then they released a new version and things got squirrelly again. This morning the system was out of service briefly.
Basically from what I can tell, Claude takes a hit when western Europe arrives for work in the morning, and then again when the U.S. east coast clocks in for the day. I got ChatGPT for some other stuff and it has never shown signs of instability, but Claude Desktop is $20/month for Pro, while the ChatGPT that supports MCP servers starts at $100/month.
There is an enormous AI bubble right now, akin to the Space Race, and it can’t last. There is a good chance that the level of service I use legitimately costs $100/month. The LLM providers have been burning venture capital in a race for market share and there’s no plausible way they’re going to make a return. We are truly wrapped up in a battle to see who gets to Artificial Super Intelligence first.
My view on this is simple; I’m doing a bit of research to find out how much of a local LLM I need in order to work. Worst case a 64GB Mac Mini is about $2k and a 96GB Mac Studio is $4k. Linux/Windows don’t cut it at this point due to lack of unified memory, you have to spend the price of the Mac Studio to get a pair of 32GB Nvidia cards, and then buy a PC on top of that. This is the doomsday scenario, I don’t expect the providers to all vanish, but there will be consolidating and rational pricing based on profit seeking rather than market share expansion.
There might be a client privacy issue here - who within the LLM providers can see the requests? Are those requests discoverable? We may see a situation where law firms MUST have an on site LLM in order to retain attorney client privilege.
Conclusion:
Over the years I have run a variety of software as a service operations that paid the rent. This is the first time I’ve just put something together, polished it a bit, and tossed it to the ravening masses. To be clear, I am NOT slick enough with either Github or Python to do this - I’m a legit FOSS developer for the first time based on the strength of Claude Code.
How many of you are using AI? There are a couple of you I chat with and I know that Claude Desktop is starting to make inroads, but I’m curious about general use.
