SLM Pipeline - Local-first HR Triage System

A local-first agentic pipeline using Ollama small language models for HR report management. This exploratory implementation uses multiple specialized models for intent detection, planning, and tool execution.

Overview

This pipeline demonstrates a multi-model approach to automated HR report management:

• all-minilm → Intent detection using embeddings
• phi3:mini → Planning and reasoning
• functiongemma → Tool execution with function calling

The system receives confidential employee reports, classifies them, generates a remediation plan, and executes actions through simulated function calls.

Features

• Intent Detection: Classifies input as harassment, burnout, policy violation, or noise
• Automated Planning: Generates step-by-step HR response plans
• Function Execution: Calls appropriate tools based on the plan
• Available Actions:
  • open_hr_case(employee_id, category, risk_level) - Open a confidential HR case
  • notify_legal(case_id) - Notify the legal department about sensitive cases
  • schedule_hr_meeting(employee_id, urgency) - Schedule follow-up HR meetings

Prerequisites

• Python 3.8 or higher
• Ollama installed and running

Installation

1. Set up Python Virtual Environment

# Create virtual environment
python -m venv venv

# Activate virtual environment
# On Windows:
venv\Scripts\activate
# On macOS/Linux:
source venv/bin/activate

2. Install Python Dependencies

pip install -r requirements.txt

3. Install Ollama Models

Make sure Ollama is installed and running, then pull the required models:

ollama pull all-minilm
ollama pull phi3:mini
ollama pull functiongemma

Verify installation:

ollama ls

Expected output:

NAME                    ID              SIZE      MODIFIED   
all-minilm:latest       1b226e2802db    45 MB     ...    
phi3:mini               4f2222927938    2.2 GB    ...    
functiongemma:latest    7c19b650567a    300 MB    ...

Usage

Run the pipeline:

python slm_pipeline.py

The default test scenario simulates an HR report:

"My Employee ID is 12345: I have been working excessive hours for months. My manager threatens retaliation if I raise concerns, and I am feeling mentally exhausted and unsafe."

Example Output

[INIT] Loading intent embeddings (MiniLM)...
==============================
[INPUT] My Employee ID is 12345: I have been working excessive hours for months. My manager threatens retaliation if I raise concerns, and I am feeling mentally exhausted and unsafe.
[INTENT] burnout
[PLAN]
[
  {
    "action": "open_hr_case",
    "employee_id": "12345",
    "category": "burnout",
    "risk_level": "HIGH"
  },
  {
    "action": "schedule_hr_meeting",
    "employee_id": "12345",
    "urgency": "HIGH"
  }
]
[STEP]
 {'action': 'open_hr_case', 'employee_id': '12345', 'category': 'burnout', 'risk_level': 'HIGH'}
[ACTION] Open HR case for employee=12345, category=burnout, risk=HIGH
...
[TOTAL LATENCY] X.XXs

Project Structure

slm/
├── slm_pipeline.py      # Main pipeline implementation
├── requirements.txt     # Python dependencies
├── models/              # (Model storage directory)
└── README.md           # This file

Architecture

1. Intent Detection: Uses MiniLM embeddings to classify input via cosine similarity
2. Planning: Phi-3 Mini generates structured JSON plans with ordered steps
3. Execution: Function Gemma selects and calls appropriate functions based on plan steps

Notes

• This is an exploratory implementation for demonstration purposes
• Function calls are simulated (print statements) - replace with actual implementations
• Function Gemma may occasionally produce non-JSON outputs; error handling is included
• All models run locally via Ollama for privacy and offline capability

Customization

Adding New Functions

1. Define the function in slm_pipeline.py
2. Add to the FUNCTIONS dictionary
3. Include in the planner's available actions
4. Add to Function Gemma's tools list

Modifying Test Scenarios

Edit the test_alert variable in the __main__ block:

if __name__ == "__main__":
    test_alert = "Your custom HR report here"
    handle_input(test_alert)

Dependencies

• ollama>=0.1.0 - Ollama Python client
• numpy>=1.24.0 - Numerical operations for embeddings
• json-repair>=0.25.0 - Robust JSON parsing

License

This is an exploratory project for demonstration purposes.

Troubleshooting

Models not found: Ensure all three models are installed via ollama pull

Ollama not running: Start Ollama service before running the pipeline

JSON parsing errors: Function Gemma may produce non-JSON output; adjust temperature or retry logic as needed