EnvBert-Agent

EnvBert-Agent is an agentic AI pipeline for environmental due diligence text classification.

It combines:

• EnvBert (domain-specific transformer backbone)
• Optional LLM fallback
• Agentic routing & confidence arbitration
• Quality control & explainability
• Modular LangGraph orchestration

Features

• Environmental domain classification using EnvBert
• Confidence-based fallback to LLM
• Agentic workflow orchestration via LangGraph
• CLI interface for quick usage
• Python SDK interface for integration
• Designed for due diligence, remediation, and compliance workflows

---

Requirements

• Python 3.9+
• Transformers (version constrained for TensorFlow compatibility)
• TensorFlow (required by EnvBert backbone)
• Ollama

Ollama Setup

1. Download and install from: https://ollama.com/download

2. Pull a model

ollama pull llama3

3. Common Error

ConnectionRefusedError: localhost:11434

It means Ollama is not running.

4. Fix:

ollama serve

By default, it runs at:

http://localhost:11434

Keep this running in a separate terminal.

Using Azure OpenAI

Set environment variables:

Windows:

set AZURE_OPENAI_API_KEY=...
set AZURE_OPENAI_ENDPOINT=...
set AZURE_OPENAI_DEPLOYMENT_NAME=...

macOS/Linux:

export AZURE_OPENAI_API_KEY=...
export AZURE_OPENAI_ENDPOINT=...
export AZURE_OPENAI_DEPLOYMENT_NAME=...

Installation

Install from PyPI:

pip install envbert-agent
ollama serve
envbert-agent "BEHP was detected in groundwater.

Python Usage

from envbert_agent import run

text = "BEHP was detected in groundwater"
result = run(text)

print(result)

from envbert_agent.config import LLMConfig
config = LLMConfig(provider="azure")

result = run(
    "BEHP was detected in groundwater",
    llm_config=config
)

print(result)

CLI Usage

After installation:

envbert-agent "BEHP was detected in groundwater"
envbert-agent "BEHP was detected in groundwater" --provider azure

Example output:

[INPUT]
raw_text: BEHP was the only SVOC detected in groundwater at concentrations exceeding the MCL.
clean_text: BEHP was the only SVOC detected in groundwater at concentrations exceeding the MCL.
language: en
quality_score: 1.0

[CLASSIFICATION]
envbert_label: Remediation Standards
envbert_confidence: 0.437928160341927
route: llm
llm_label: Contaminants
llm_confidence: 0.9
final_label: Contaminants
final_confidence: 0.9

[META]
llm_reasoning: The text mentions a specific contaminant (BEHP) and its concentration exceeding the Maximum Contaminated Level (MCL), indicating the presence of contaminants in groundwater.
decision_trace: LLM fallback
key_phrases: ['behp', 'svoc', 'detected', 'groundwater', 'concentrations', 'exceeding']

[MONITORING]
drift_flag: False

[INPUT]
raw_text: soil and groundwater are both contaminated on the site
clean_text: soil and groundwater are both contaminated on the site
language: en
quality_score: 1.0

[CLASSIFICATION]
envbert_label: Contaminated media
envbert_confidence: 0.7678613004581079
route: accept
llm_label: N/A
llm_confidence: N/A
final_label: Contaminated media
final_confidence: 0.7678613004581079

[META]
llm_reasoning: N/A
decision_trace: EnvBert backbone
key_phrases: ['soil', 'groundwater', 'both', 'contaminated', 'site']

[MONITORING]
drift_flag: False

Direct CLI Invocation from Python

from envbert_agent.cli import main

main(["BEHP was detected in groundwater"])

Architecture: Graph Edges & Flow

START
  ↓
[preprocess] ────────────────┐
  ↓                           │
[envbert] ────────────────────┤
  ↓                           │
[arbitrate]                   │  Conditional Router:
  ├─ (quality < 0.4)          │      route = "review"
  ├─ (confidence ≥ 0.75) ─────┼──────→ route = "accept"
  └─ (confidence < 0.75) ─────┘      route = "llm"
       ↓
    ┌──────────┬──────────┬──────────┐
    ↓          ↓          ↓
  (review)   (accept)   (llm)
    ↓          ↓          ↓
    │ ─────────→[llm]←─────
            ↓
        [evaluate]
            ↓
        [explain]
            ↓
        [monitor]
            ↓
           END

License

MIT License

See the LICENSE file for details.

⚠️ Notes

• This package depends on the EnvBert backbone.

• Transformers version is constrained for TensorFlow compatibility.

• Future versions may migrate to a PyTorch backend for improved compatibility and lighter installation footprint.