retrieval

Inter-Modal Retrieval

Local-first Codex session archive, native desktop, smart backup, and read-only MCP tool.

Code for the paper "Evaluating cross-lingual textual similarity on dictionary alignment problem"

Multimodal RAG system matching field photos + damage reports to replacement parts (FastAPI, Cerebras vision LLM, sentence-transformers)

Local corpus retrieval over your own text: BM25 + SPLADE v3 + ColBERTv2, with synthetic eval and an optional relevance judge

Retrieval over a large engineering corpus for Heliosnet (LlamaIndex and Qdrant), with retrieval quality enforced as a CI gate.

Your RAG is only as good as how you cut up the documents. Measured whether the '80% vs 50%' headline generalizes to a second document: the margin nearly vanishes, and a retriever bug flips which strategy wins.

This repository is designed for absolute beginners who want to master RAG from the ground up in just 10 days

Build your own second brain with supermemory. Extremely fast, scalable, memory API for the AI era.

Own integration of PokéAPI

Interactive AI architecture console demonstrating layered cognitive system design, adaptive routing, and cost-aware execution modeling.

A domain-specific conversational AI assistant based on a custom Hybrid RAG architecture, designed to answer NLP & LLM UNISA course questions using local documents and on-premise LLMs via Ollama.

The smallest useful RAG, in 40 lines — a Markdown folder, sqlite-vec, and a CLI. Companion to the 'What is RAG?' primer at cloudandsre.com.

Precise long-term memory for LLM agents on PostgreSQL: documents stored as atoms, answers always cited. Hermes skill (schema + scripts + deployment).

HNSW and IVF-PQ parameter sweeps producing real recall-vs-latency curves, the filtered-ANN problem demonstrated, and a measured answer to whether re-ranking is worth its latency.

bleach the rot away

Multi-tenant document retrieval on plain PostgreSQL — full-text search, trigram similarity and unaccented matching, with no vector database.

RAG that goes past embed-and-pray. BM25 + semantic fusion with reranking cut retrieval errors 81%, and every claim gets checked against its citation before it ships.

Agent-readable memory for markdown notes: enforced schema, validator, BM25 + graph retrieval, MCP server, and an eval that tells you the truth