Back to projects

AI Tools & Applications

MemoryLens — AI Semantic Search Engine

Find a movie, book, song or game from a half-remembered fragment.

top-1 accuracy, full pipeline
92%top-1 accuracy, full pipeline
top-1, retrieval only
54%top-1, retrieval only
benchmark queries
48benchmark queries
categories
6categories

// at a glance

The problem and the idea

The problem

Everyone has had a title on the tip of their tongue. Normal search needs keywords you no longer remember, and chatbots often invent titles that don't exist.

The idea

You pick a category and describe what you remember in plain language. MemoryLens searches a real catalog with hybrid retrieval, reranks the candidates and lets an LLM reason over the shortlist. Each match comes with a confidence score and an explanation. When the catalog has no confident match, it names the real title from model knowledge and labels it as such.

// who it's for

Who it's for

  • 01

    People trying to recall a film, series, book, song, game or actor

  • 02

    Readers and gamers who remember a scene or a feeling, not a name

  • 03

    Anyone tired of chatbots confidently inventing titles

// how it works

How it works

queryunderstandhybrid retrievalpgvectorcross-encoderLLM reasoningconfidencetop matchrecall@1: 54% retrieval only → 92% full pipeline

1/5 Understand the query

Swipe sideways to see the whole diagram →

// screenshots

Screenshots

// key features

Key features

  • Six categories: movies, TV series, songs, books, games and actors
  • Grounded in a real catalog, so titles are never made up
  • A soft "switch category?" hint when a query clearly belongs elsewhere
  • Accounts with search history, per-user rate limits and a daily quota
  • Local embeddings and reranking; the LLM costs about $0.0003 per search
  • Production setup with nginx, Redis caching, security headers and CI

// design decisions

Design decisions

  1. 01

    Every stage was measured on a fixed 48-query benchmark instead of guessing which part helps.

  2. 02

    Retrieval mostly improves depth (recall@5); the LLM layer is what moves the top-1 answer.

  3. 03

    Embeddings and reranking run locally, so only the final reasoning step costs money.

// built with

Built with

  • FastAPI
  • React
  • PostgreSQL
  • pgvector
  • Redis
  • sentence-transformers
  • Docker

Run it locally

$ docker compose up --build