Applied AI & knowledge systems
ArchiveAI
Turn unstructured files into a source-cited knowledge workspace.
- Status
- Working open-source product
- My role
- Solo product, AI, backend, and interface engineering
- Timeframe
- March - June 2026
- Core technology
- Next.js · FastAPI · LangGraph · Docling
01 / Context
The problem
Important knowledge is trapped across documents in incompatible formats. Generic chat tools make it difficult to inspect what was indexed or trace an answer back to its evidence.
02 / Product
The implemented solution
ArchiveAI gives users a complete document workspace: upload and inspect files, search semantically, manage conversations, and receive streamed answers that remain linked to supporting sources.
03 / System
Architecture, in plain language
A Next.js interface communicates with FastAPI services for ingestion and retrieval. Docling normalizes document structure, PostgreSQL and pgvector store searchable chunks, and a LangGraph workflow orchestrates grounded responses.
04 / Judgment
Important decisions
- Designed the product around inspectable sources instead of treating citations as an afterthought.
- Separated ingestion, retrieval, and conversation concerns so each can evolve independently.
- Used persistent conversations to support research sessions rather than one-off prompts.
05 / Boundaries
Limitations and next work
Answer quality still depends on source quality, parsing fidelity, and retrieval coverage. Generated responses require human verification for high-stakes use.
06 / Implementation
Focused technology
Next.js · FastAPI · LangGraph · Docling · PostgreSQL · pgvector