memini-vision 0.1.0

Shared vision library for image storage, CLIP embedding, and cross-modal search

memini-vision

Shared vision library used by both videre-mcp (the MCP server that exposes vision tools to agents) and memini-ai-dev (the semantic memory server that fuses image results into its existing RRF). Owns image storage, CLIP embedding, the memories_image table CRUD, and cross-modal query logic.

No FastMCP, no CLI, no HTTP server. Library only — the two consuming servers register their own tools.

Install

pip install memini-vision

Environment variables

Variable	Default	Description
MEMINI_IMAGE_SEARCH_ENABLED	false	Master kill-switch (consumer checks flag)
MEMINI_IMAGE_CLIP_MODEL	clip-ViT-B-32	clip-ViT-B-32 or clip-ViT-L-14
MEMINI_IMAGE_CLIP_DEVICE	auto	auto, cpu, or cuda
MEMINI_IMAGE_DIR	~/.memini-ai/images	Filesystem image storage root
MEMINI_IMAGE_MAX_FILE_SIZE	10485760	Max image file size in bytes (10MB)
MEMINI_IMAGE_DB_URL	(empty)	PostgreSQL connection string (falls back to MEMINI_DB_URL)

Quickstart

from memini_vision import VisionConfig, ImageStore, ClipEmbedder, ImageIndex, ImageQuery

cfg = VisionConfig()                     # reads MEMINI_IMAGE_* env vars
store = ImageStore(cfg.image_dir)        # filesystem sharded storage
embedder = ClipEmbedder(cfg.clip_model, cfg.clip_device)
index = ImageIndex(cfg.db_url)           # asyncpg pool, memories_image CRUD
query = ImageQuery(embedder, index)       # cross-modal CLIP search

# Encode + store an image (lazy model load on first call)
vec = embedder.encode_image(pil_image)    # list[float], zero-padded to 768 for B/32
record = await index.insert(memory_id, sha256, vec, file_path, mime_type)
results = await query.search_by_text("terminal traceback", limit=5)