Czar Brain / Hermes Retrieval Infrastructure

System Summary

Built a self-hosted knowledge retrieval system that makes Obsidian notes, project documents, and approved files searchable from Hermes with source citations.

Business Problem

Hermes and other agents could not reliably answer questions from private knowledge because retrieval was either missing, inconsistent, or lacked source citations.

What Czar Studios Built

Hybrid retrieval pipeline combining Qdrant dense/sparse vectors, SQLite FTS5/BM25, reciprocal-rank fusion, chunking rules, citation formatting, secret exclusions, and a recurring reindex timer.

Architecture and Data Flow

Obsidian Markdown files are parsed and chunked; dense and sparse embeddings are stored in Qdrant; full-text indexes live in SQLite; a fusion layer ranks results and returns citations. Reindexing runs on a schedule so new notes are searchable without manual steps.

Integrations

Obsidian, Qdrant, SQLite, Hermes, Python indexing service, and approved file sources.

Permissions, Guardrails, and Human Review

Secret exclusions prevent sensitive paths from being indexed; source citations make answers traceable; reindexing is scheduled and logged.

Testing and Verification

Verified retrieval accuracy on known questions, citation correctness, reindex completion, and exclusion of secret paths.

Practical Outcome