Back to Projects
Case Study Stub2025

Debo — Engineering a Personal RAG & Agentic Memory Architecture

Building a personal memory ingestion engine that embeds notes, web clips, audio, and conversations for sub-second agentic recall.

Vector RAG

Embedding & Retrieval

Multimodal

Text, Audio & Web Clips

Solo Build

End-to-End Pipeline

1. The Core Problem

Knowledge workers capture fragments across dozens of platforms daily — browser tabs, voice memos, quick notes, and PDF papers. Standard search engines require keyword recall; LLMs lack personal episodic memory.

Debo was built as an autonomous personal memory pipeline: continuously chunking, embedding, indexing, and surfacing semantic context during conversation.

TODO(shaswat): What vector database and embedding model were chosen for Debo (e.g. Cloudflare Vectorize, Pinecone, OpenAI text-embedding-3)?

2. Schema & Data Model Decisions

Memory documents were structured with hierarchical metadata tags, chunk offsets, and temporal decay weights:

TODO(shaswat): Provide the memory chunk data model and how metadata filtering is combined with cosine similarity search.

3. Hardest Scaling or Correctness Problem

Balancing chunk size precision vs retrieval recall, and preventing hallucinations when query concepts span multiple disjoint notes:

TODO(shaswat): Detail the exact retrieval reranking or prompt synthesis challenges you solved in Debo.

4. One Tradeoff I'd Reverse

Tradeoffs made between cloud vector storage and local-first embedded SQLite/Wasm storage:

TODO(shaswat): What architectural or pipeline tradeoff in Debo would you reverse today?