ReeCall.ai — Checkpoint (v1)

ReeCall.ai is a lightweight (~50K parameters) Graph Neural Network (ReeCall / ReeCallGNN) that learns file-level dependency embeddings for TypeScript/JavaScript codebases and retrieves relevant engineering context as structured JSON.

Model Properties

Property	Detail
Architecture	MLP Encoder + 1-Hop Neighbor Mean Aggregation + Projection Head
Parameters	~50,000 (CPU-friendly)
Input Feature Dim	128 (structural metrics, language, file roles, symbol & path hashes)
Embedding Dim	32 (L2-normalized for cosine similarity)
Training Objective	Co-import prediction via Contrastive Loss (margin=0.5)

Quick Usage

import torch
from reecall_ai import ReeCall

# Initialize ReeCall model
model = ReeCall(input_dim=128, hidden_dim=64, embed_dim=32)

# Load checkpoint weights
checkpoint = torch.load("model.pt", map_location="cpu", weights_only=True)
model.load_state_dict(checkpoint)
model.eval()

print("ReeCall model loaded successfully!")

Files in Repository

• model.pt: Trained PyTorch ReeCall model state dict (~200KB).
• file_embeddings.npy: Example 32-dim normalized file embeddings matrix.
• default.ini: Model & training hyperparameters config.

Repository & Documentation

• GitHub Repository: https://github.com/om-ghante/reecall.ai
• License: MIT