Cosolvent
Cosolvent is an open-source (MIT License) platform designed to simplify assembly and access to AI-powered search and organization tools. Built specifically for “thin markets”—where participants are scattered and struggle to find each other … and then talk about complex stuff … Cosolvent provides a foundation for anyone wanting to explore LLM+RAG automation and build specialized search and exchange environments. The high level concept is in the following diagram:
Cosolvent is not a universal solution to automate thin markets. That is unrealistic, if not impossible. Cosolvent is designed to be a general purpose “empty vessel” that can integrate a variety of AI-related tools and provide generalized support for their data management and operations. If suitable tools are selected and integrated into Cosolvent, the amalgam can be adapted to a wide range of thin markets. When ClientSynth generates a representative set of fake, but realistic, users of the various required types, the pieces combine to form a realistic thin market automation simulator like GPSim that can be used to experiment and evangelize for a new thin market solution. Some of the key generic support functions include:
- Asset Onboarding and Management: Users can upload documents, images, links, videos, etc. to their private folder. The assets will be held in pending status (downloadable, but unavailable to anyone but the user and invisible to any of the AI functions … until the user decides how, where and when each one should be made available.
- Microservice Architecture: Cosolvent is based on an API-based microservice architecture that uses Python for the backend and Node.js for the frontend. This allows Cosolvent to connect to any service (local, cloud, paid or free) that offers a suitable API. Administrators can direct different functions to different LLMs and swap out the chosen LLMs as new and better ones appear.
- Prompt Management and Execution: Cosolvent has an AI “playground” where administrators can experiment with new prompts and links to new AI assets (LLMs, embedding engines, etc.) … without having to code.
- Language Translations: The internal language of record for Cosolvent is English. Cosolvent can apply an LLM to translate uploaded materials from a number of languages into English for use in AI operations. Similarly, Cosolvent can translate output data presentations from English to a supported language at the user’s discretion. In the short run, this process will likely result in some less-than-perfect translationis, but the trends in AI development suggest that it will improve very quickly.
- Intelligent Profile Creation: Users can build profiles by uploading existing documents—specifications, certifications, product descriptions, or any relevant materials. Cosolvent’s AI automatically extracts and structures this information into searchable, matchable profiles with a consistent look and feel without requiring users to fill out complex forms or learn new systems.
- Curated Industry Context: Each market implementation includes a carefully assembled library of industry-specific information—standards, regulations, terminology, and best practices—that provides essential context for accurate matching and relevant responses.
Building Toward a Thin Market Ecosystem
When the Cosolvent toolkit is built out and customized for a given thin market, the system sponsor must supply a lot more information than Deeperpoint is putting into the open source. That will require design, marketing and support to appeal to the specific people and organizations that are expected to participate in the market. The sponsor will also have to assemble and curate real-world documents, links and assets to create a trustworthy industry context reference. In most instances, the participants won’t be aware of Cosolvent’s design or operations. They will just be ordinary people that are persuaded to use the resulting system. Recruiting and supporting the market participants will be a separate (and probably daunting) task for the thin market system sponsors.
That’s where Deeperpoint’s companion GPSim simulator and ClientSynth projects can help. ClientSynth uses AI to generate a synthetic, but realistic, ecosystem of users and participants for a chosen thin market scenario. Then, Cosolvent hosts and organizes the system of AI tools (LLMs, RAG, prompts and presentations) that are chosen to handle the challenges of each market. That creates an initial playground where developers, and soon registered users, can play with realistic situations and scenarios.
Conceptually, the simulator must be customized for each specific thin market. Then, a market sponsor must hire and manage a team of people to recruit and support the many participants and end users. The challenge for the Cosolvent design process is that we couldn’t see how to build and test a fully working implementation without first organizing a corresponding ecosystem. The chicken-and-egg problem again. Cosolvent plus the simulator is our attempt to break that cycle.
Visitors are welcome to poke around the Cosolvent repository on Github. You can clone a current version and follow the documented guidelines to set up a docker instance and play with your own version. Since it is released under the MIT license, you are welcome to fork the design and go off on your own development path. We hope that you will keep in touch and perhaps give back to the project, but that is totally up to you.