Skip to content

all-MiniLM-L6-v2 Adapter

Encodes text as 384-dimensional dense vectors for semantic similarity, search, and clustering.

Model details

Field Value
Model sentence-transformers/all-MiniLM-L6-v2
Task embed
Domain general
License Apache 2.0

Install

pip install synapse-adapter-sdk
pip install sentence-transformers

Verified output schema

model.encode() returns a numpy.ndarray of shape (1, 384). The adapter maps this to:

  • payload.modality"embedding"
  • payload.vectorlist[float] of length 384
  • payload.vector_dim384
  • payload.embedding_model"sentence-transformers/all-MiniLM-L6-v2"

Example access:

vector = result_ir.payload.vector       # list[float], length 384
dim    = result_ir.payload.vector_dim   # 384

If the model output is missing or the wrong shape, vector=[] and vector_dim=0 are set rather than raising. Provenance confidence is fixed at 1.0.

Supported task types

  • embed

Supported domains

  • general

Usage example

import time
from sentence_transformers import SentenceTransformer
from all_minilm_adapter import AllMiniLML6V2Adapter

model   = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
adapter = AllMiniLML6V2Adapter()

# 1. Prepare model input
model_input = adapter.ingress(ir)
# {"sentences": ["The quick brown fox jumps over the lazy dog."]}

# 2. Run the model (caller's responsibility)
t0 = time.monotonic()
model_output = model.encode(model_input["sentences"])
latency_ms = int((time.monotonic() - t0) * 1000)
# numpy.ndarray shape (1, 384)

# 3. Convert output back to canonical IR
result_ir = adapter.egress(model_output, ir, latency_ms=latency_ms)

# 4. Access the embedding
vector = result_ir.payload.vector   # list[float], length 384

Source

github.com/synapse-ir/adapters