mindpadi/intent_encoder

Intent Encoder (MindPadi)

The intent_encoder is a Sentence Transformer model used in the MindPadi mental health assistant for encoding user messages into dense embeddings. These embeddings support intent classification, similarity search, and memory recall workflows. It plays a foundational role in the semantic understanding of user inputs across various MindPadi features.

🧠 Model Overview

• Architecture: Sentence-BERT (all-MiniLM-L6-v2 base)
• Task: Sentence Embedding / Semantic Similarity
• Purpose: Embed user queries for intent classification, vector search, and memory retrieval
• Size: ~80M parameters
• Files:
  • config.json
  • pytorch_model.bin or model.safetensors
  • tokenizer.json, vocab.txt
  • 1_Pooling/, 2_Normalize/ (Sentence-BERT components)

🧾 Intended Use

✔️ Primary Use Cases

• Semantic embedding of user inputs for intent recognition
• Matching new messages against known intent samples (data/processed_intents.json)
• Supporting vector similarity in MongoDB Atlas Search or ChromaDB
• Powering memory in LangGraph agentic workflows

🚫 Not Recommended For

• Direct intent classification (this model returns embeddings, not classes)
• Use outside of NLP (e.g., image, audio)

🧪 Integration in MindPadi

• app/chatbot/intent_classifier.py: Uses this model to compute sentence embeddings
• app/chatbot/intent_router.py: Leverages vector similarity for intent matching
• database/vector_search.py: Embeddings are stored or queried from MongoDB vector index
• app/utils/embedding_search.py: Embeds utterances for real-time nearest-neighbor lookup

🏋️ Training Details

• Base Model: sentence-transformers/all-MiniLM-L6-v2 (pretrained)
• Fine-tuning: Optional domain-specific contrastive learning using pairs in training/datasets/fallback_pairs.json
• Script: training/fine_tune_encoder.py (if fine-tuned)
• Tokenizer: BERT-based WordPiece tokenizer
• Max Token Length: 128

📈 Evaluation

While this model is not evaluated via classification metrics, its embedding quality was assessed through:

• Cosine similarity tests (intent embedding similarity)
• Intent clustering accuracy with KMeans in vector space
• Recall@K for correct intent retrieval
• Visualizations: UMAP plots (logs/intent_umap.png)

Results indicate:

• High-quality clustering of semantically similar intents
• ~91% Top-3 Recall for known intents

💬 Example Usage

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("mindpadi/intent_encoder")

texts = ["I want to talk to a therapist", "Book a session", "I'm feeling anxious"]
embeddings = model.encode(texts)

print(embeddings.shape)  # (3, 384)

🧪 Deployment (API Example)

import requests

endpoint = "https://api-inference.huggingface.co/models/mindpadi/intent_encoder"
headers = {"Authorization": f"Bearer <your-token>"}
payload = {"inputs": "I need help managing stress"}

response = requests.post(endpoint, json=payload, headers=headers)
embedding = response.json()

⚠️ Limitations

• English-only
• Short, clean sentences work best (not optimized for long documents)
• Does not directly return intent labels — must be paired with clustering or classification logic
• May yield ambiguous vectors for multi-intent or vague inputs

📜 License

MIT License – open for personal, academic, and commercial use with attribution.

📬 Contact

• Project: MindPadi Mental Health Assistant
• Team: MindPadi Developers
• Email: [[email protected]]
• GitHub: [https://github.com/mindpadi]

Last updated: May 2025