shivvr

Semantic memory service. Cuts text into pieces, embeds them, stores them, finds them again.

Rust + ONNX + sled. Runs in Docker on port 8080.

What it does

• Ingest — chunks text by sentence boundaries, embeds each chunk with BGE-small-en-v1.5 (384d), stores in sled
• Search — cosine similarity over organize embeddings, returns ranked chunks
• Dual embedding — local BGE-small for organize (384d), optional OpenAI ada-002 for retrieve (1536d)
• Per-agent encryption — identity-keyed orthogonal matrix encryption on embeddings
• Vec2text inversion — reconstruct approximate text from embeddings (T5-based, optional)
• Emotion tagging — vedana/feeling-tone metadata on each chunk at ingest time

Quick start

docker compose up -d

Builds from source, downloads the BGE-small ONNX model, starts on :8080.

API

Memory

Method	Endpoint	Description
GET	/health	Status, model info, session/chunk counts
GET	/memory	List all sessions
POST	/memory/:session/ingest	Ingest text (auto-chunks + embeds)
GET	/memory/:session/search?q=...	Semantic search
GET	/memory/:session/info	Session metadata
DELETE	/memory/:session	Delete a session

Crypto

Method	Endpoint	Description
POST	/agent/:id/register	Register agent encryption keys
POST	/agent/:id/encrypt	Encrypt embeddings with agent key
POST	/agent/:id/decrypt	Decrypt embeddings with agent key

Inversion

Method	Endpoint	Description
POST	/invert	Reconstruct text from embedding vector

Ingest example

curl -X POST http://localhost:8080/memory/my-session/ingest \
  -H "Content-Type: application/json" \
  -d '{"text": "The harbor was quiet at dawn. Only the sound of halyards against aluminum masts."}'

Search example

curl "http://localhost:8080/memory/my-session/search?q=morning+at+the+marina&top_k=5"

Environment

Variable	Default	Description
PORT	8080	Listen port
DATA_PATH	/data/shivvr	sled database directory
MODEL_PATH	/models/bge-small-en-v1.5.onnx	BGE-small ONNX model
TOKENIZER_PATH	/models/tokenizer.json	BGE tokenizer
OPENAI_API_KEY	—	Enables ada-002 retrieve embeddings

Stack

• Rust — axum, tokio, sled, ort (ONNX Runtime)
• Embedding — BGE-small-en-v1.5 (BAAI), 384 dimensions, L2-normalized
• Storage — sled embedded database, persistent Docker volume
• Inversion — T5-base hypothesis + corrector trained on MSMARCO (optional)

License

Proprietary. All rights reserved.