Aether
Retrieval-grounded reasoning over private corpora, with citations strong enough to be auditable.
Overview
Problem
Solution
Tech stack
- TypeScript
- Python
- PostgreSQL
- pgvector
- LLM tooling
Engineering decisions
- Evaluation harness is the unit of progress; nothing ships without a measurable retrieval or faithfulness delta.
- Citations are first-class — produced by the model, verified against the retrieved set, and shown inline.
Engineering notesWhat's shaping this build.
Design goals, philosophy, planned architecture, and where the project stands today. No fabricated benchmarks — only what's actually driving decisions.
What's shaping this build.
Design goals, philosophy, planned architecture, and where the project stands today. No fabricated benchmarks — only what's actually driving decisions.
- Citations that are verifiable, not decorative.
- Evaluation that catches regressions before deploy.
- Reproducible indexes so answers are attributable to a corpus version.
If you can't measure it, don't ship it. Retrieval quality and answer faithfulness both need offline scores that gate releases.
Hybrid sparse+dense retrieval over chunked, versioned documents in pgvector; a reasoning layer producing chunk-level citations; an offline eval harness scored against curated question sets.
Retrieval pipeline and eval harness scaffolded; iterating on chunking strategy and citation verification.
Links
Repository and demo links will appear here once the project is ready for public review.