Semantic search engine for local files (text and images). It can index files in directory (recursively) and allows searching via a text query or/and image query. Includes a Gradio web UI for interaction.

Microservice architecture:

• file-discovery-service watches the /data directory, pushes file changes to a Redis queue
• main-service consumes directory updates, uses embedding microservice to generate file embeddings, and stores the results in a Qdrant database for Cosine search
• embedding-service hosts CLIP-like model for calculating embeddings from texts and images

Model can be changed in global_config.env. Need to specify model checkpoint (from HuggingFace), embedding size and target model image resolution. Supports models:

• openai/clip-vit-large-patch14
• jinaai/jina-clip-v1
• jinaai/jina-clip-v2
• google/siglip-base-patch16-224 (and other SigLip models) - currently useless
• google/siglip2-base-patch16-224 (and other SigLip 2 models) - currently useless

Example default config for jinaai/jina-clip-v1:

EMBEDDING_MODEL_CHECKPOINT = jinaai/jina-clip-v1
EMBEDDING_SIZE = 768
TARGET_IMAGE_SIZE = 224,224

docker compose up -d --build

Web UI: http://localhost:8002/ui Main API docs: http://localhost:8002/docs

POST /api/v1/search

Performs a search. Takes optional text_query (form), image_query (file), and top_n (form) as input. Returns filenames and scores

POST /api/v1/files

Uploads one or more files to be indexed. Takes a list of multipart file uploads

POST /api/v1/index

Triggers indexing for a list of file paths already existing in /data directory

DELETE /api/v1/index

Removes files from the search index based on a list of file paths. File is not deleted. It will be reindexed on next app restart