memelo

A cloud-native meme management system with multi-modal search. Upload images via Telegram, extract text with OCR, generate semantic embeddings, and search your collection by text, fuzzy match, or meaning.

How to start

In repo root folder:

cp .env.example .env

# edit .env - must have valid GOOGLE_CREDS_PATH and IMAGE_EMBEDDING_PROJECTNAME

docker compose up -d

How to obtain google creds?

Google docs: https://docs.cloud.google.com/docs/authentication/application-default-credentials
App need a json file from google cloud console, if you logged in gloud cli - it located at $HOME/.config/gcloud/application_default_credentials.json

So, minimal instruction:

• Register at google cloud
• Enable two apis - Cloud Vision API and Vertex AI API. (I don't remember where is enable buttons exactly)
• Either:
• • Create service account
• • Download key for it
• Or:
• • Install gcloud cli tools
• • Login to your google account. It will create key automatically on your PC in default location.

Anyway, check documentation about ADC, there is a lot of ways to pass credentials to application

Architecture

Integration services  ──(gRPC)─►  storage-service
                                       │
                             ┌─────────┼─────────┐
                             ▼         ▼         ▼
                        MinIO S3  Elasticsearch  Google Cloud
                       (images)   (metadata +    (Vision OCR +
                                  vectors)       Vertex AI embeddings)

Modules:

Module	Description
storage-service	Core service — image processing, OCR, embeddings, search, export
telegram-service	Telegram bot frontend — upload, search, delete via chat/inline
common	Shared config, logging, and helper utilities
gen	Generated protobuf/Connect RPC code (do not edit)
proto	Protocol buffer source definitions

Features

• Image processing — converts uploads to JPEG, generates thumbnails (libvips)

• OCR — extracts text from images via Google Cloud Vision API

• Semantic embeddings — 1408-dimensional vectors via Google Vertex AI (multimodalembedding@001)

• Deduplication — hash-based and embedding similarity checks on upload to avoid image duplicates

• Multi-modal search pipeline — ordered searchers, first match wins:

  Order	Searcher	Strategy
  0	SimpleSearcher	Full-text on OCR result
  10	IdSearcher	Direct UUID lookup by image id
  20	FuzzySearcher	Fuzzy text match
  30	TextEmbeddingSearcher	Semantic vector search
  40	AllSearcher	List all (empty query fallback)

Tech Stack

• Language: Go 1.24
• API: Connect RPC (gRPC over HTTP/2)
• Search: Elasticsearch 8.16
• Object storage: MinIO (S3-compatible)
• User data: MongoDB
• Image processing: bimg (libvips wrapper) Require libvips to be installed
• OCR: Google Cloud Vision API
• Embeddings: Google Vertex AI
• DI: Uber fx

Local Development

Start dependencies:

 docker compose up elasticsearch minio -d

This starts Elasticsearch (:9200), MinIO (:9000).

Build and run storage-service:

cd storage-service
cp .env.example .env

# edit .env here
# app must have valid IMAGE_EMBEDDING_PROJECTNAME (google console project id like word-word-111111-a1)

go run .

Prerequisites for storage-service: libvips must be installed (vips-dev / vips).

Docker:

docker build -f Dockerfile-storage-service -t memelo-storage .
docker build -f Dockerfile-telegram-service -t memelo-telegram .