VisMem RAG Fine-tuned: gemma-3-270m-it-vismem-rag-0310-1943-vismem-rag-0311-1416-vismem-rag-0313-1351

Inference

import requests
from vismem_core import VisMem
from sentence_transformers import SentenceTransformer
from transformers import pipeline

# 1. Load VisMem
data = requests.get(
    "https://huggingface.co/datasets/broadfield-dev/gemma-3-270m-it-vismem-rag-0310-1943-vismem-rag-0311-1416-vismem-kb-0313-1351/resolve/main/vismem.png",
    headers={"Authorization":"Bearer <HF_TOKEN>"}).content
mem  = VisMem.from_png_bytes(data)
emb  = SentenceTransformer('all-MiniLM-L6-v2')

# 2. RAG query
q_vec   = emb.encode([your_question])[0]
results = mem.search(q_vec, k=3)
context = "\n---\n".join(results)

# 3. Prompt
system = (
    "You are a helpful AI Assistant with visual memory.\n"
    "### RAG MEMORY (Vector Database):\n"
    "[Uploaded Doc]: None\n"
    f"[Knowledge Base]: {context}\n"
    "### EPISODIC MEMORY (Past Chat):\n"
    "[History]: None"
)

pipe = pipeline("text-generation", model="broadfield-dev/gemma-3-270m-it-vismem-rag-0310-1943-vismem-rag-0311-1416-vismem-rag-0313-1351")
print(pipe([
    {"role":"system","content":system},
    {"role":"user","content":your_question}
], max_new_tokens=200))

Config

{ "dataset_name": "openai/gsm8k", "rag_columns": [ "answer" ], "question_col": "question", "answer_col": "answer", "split": "train", "data_config": "main", "total_kb_docs": 7473, "vismem_dim": 384, "vismem_width": 4352, "vismem_height": 4352, "kb_repo": "broadfield-dev/gemma-3-270m-it-vismem-rag-0310-1943-vismem-rag-0311-1416-vismem-kb-0313-1351" }