RAG that remembers when, not just what

What it does

Utopia ingests documents—PDFs, spreadsheets, web pages, RSS feeds—and extracts entities and facts into a knowledge graph backed by PostgreSQL. Every fact carries a validity interval (valid_from / valid_to) and a pointer back to the source sentence, so the graph can be queried as-of any historical moment. It answers questions via hybrid retrieval (full-text Tantivy plus pgvector embeddings) and streams cited responses through any OpenAI-compatible endpoint.

The interesting bit

Most knowledge bases overwrite corrections; Utopia treats facts as append-only, closing the previous version rather than erasing it. A client-side timeline lets you drag a slider and redraw the graph neighborhood as it existed at that exact moment—though server-side as-of queries are still on the roadmap.

Key highlights

• One Rust binary plus Postgres: no Elasticsearch, no dedicated vector service, no message queue.
• Hybrid search fuses Tantivy full-text and pgvector with reciprocal rank fusion, including jieba tokenization for Chinese.
• LLM-extracted facts land in a review queue for human confirmation; entity merges are reversible and logged.
• Chat can query registered Postgres databases alongside your documents.
• Runs fully offline with local models via Ollama or vLLM.

Caveats

• Credentials (LLM keys and DB connection strings) are stored unencrypted in Postgres; the README explicitly warns to run it on a trusted network.
• At v0.1 with a single maintainer, schema changes between releases and there is no guaranteed upgrade path.
• Server-side temporal queries and the Datalog reasoning engine are not yet implemented.

Verdict

Worth a look if you need an air-gappable, self-hosted knowledge base for compliance or historical audit trails. Skip it if you want a mature SaaS with encryption-at-rest and guaranteed migrations.