Antonio Escohotado's legacy, explorable with AI
How do you turn a philosopher's work into something you can ask? Antonio Escohotado wrote for decades, and his legacy was scattered across books, articles, videos and audio recordings. The project's question was easy to pose and hard to solve: how to make all of it searchable, readable and open to questions without betraying it.
I built the Portal de Antonio Escohotado in collaboration with Jorge Escohotado: a platform that digitizes, indexes and connects his work.
What has been built
- 21 complete books digitized, plus 213 articles, 296 videos, 109 audio recordings and 152 PDFs from his library.
- Faceted search across all content, with taxonomy filters.
- El Oráculo: a chat that answers questions about the work citing its sources.
- Registration and subscription with a freemium paywall, plus Telegram, WhatsApp and Discord bots to talk with the agent.
Hybrid search over thousands of fragments
A self-hosted Typesense instance with nine collections combines lexical and semantic search —OpenAI embeddings generated at indexing time— over roughly ten thousand fragments. Books are chunked by chapter, keeping each chapter's title in the fragment, and the 152 PDFs go through an asynchronous parsing pipeline with LlamaParse.
The hard part: keeping it from hallucinating
A RAG system built on a philosopher cannot afford to make things up. A fabricated quote is not a minor bug: it is a betrayal of the author. Four decisions keep it in check:
- prompts that require citing the source or staying silent.
- Synthesis tools that verify every quote on the server against the retrieved fragments.
- A search guide that teaches the agent to search by concept, not by headline.
- A
LiteLLMgateway with a curated model catalog and per-subscription-tier limits.
The heart: Payload CMS
Everything in the portal goes through Payload CMS: the articles, the books, the quotes, the videos, the audio recordings, the PDFs, the taxonomies and the users. The data model is code and, on top of that model, Payload generates the API and the admin panel the editorial team works with every day.
That provides two things. First, pipeline traceability: every document shows its indexing status in the admin —synced or not indexed—, so it is always clear which part of the content is available to search and to the agent. Second, extensibility: every change triggers the hooks that keep the index in sync, with no manual processes.
And most importantly for us: agents are content too. On top of Payload we built PayloadAgents, our open source project that adds agents and search profiles as CMS collections. El Oráculo is not hard-coded: it is configured as data —its prompt, its tools, its model— and operated from the same panel as the rest of the content.
The origin of the platform
There is a timeline I like to tell: Escohotado was not built on top of the platform —the platform was built from Escohotado—. This was the pioneering project from which the stack that is now Zetesis Portal was standardized, and in 2026 the portal itself migrated to that standard.
Self-hosted infrastructure, no Kubernetes
The entire platform runs on a single 6 GB RAM VPS —the lowest possible cost—: Keycloak for identity, PostgreSQL, the RAG index with Typesense and Payload CMS, each in its own container behind a reverse proxy. No Kubernetes: at this scale, well-organized Docker Compose is cheaper to operate and to understand.
- CI/CD with
doco-cd: every push to the repository triggers deployment of the updated stack. - Automatic backups of the database and the content.
- Images signed with
cosignand automated releases with release-please.
The complete system —search, agent, paywall and bots— is already running on that minimal server, still in pre-production: the public launch is underway.
Conclusions
Any organization with a documentary legacy —a publisher, a foundation, a company with scattered knowledge— can turn it into something searchable and askable. The technology exists; the hard part is rigor: ensuring every answer can point exactly to where it comes from.
More real-world cases
Zetesis Portal and PayloadAgents: how we built our SaaS
Konect: audio transcription and analysis
Applied AI consulting for business
Classifying documents into 77 categories without training a model
From standard to graph to project
