Message-Analyzer

Message Analyzer is a forensic tool that analyzes text conversations to extract key criminal elements, with a special focus on murder-related analysis for now. The system uses Ollama's Mistral 7B model and is deployed via a structured Flask-ML API.

The project includes:

• Conversation-Based Crime Analysis
• Criminal Element Extraction
• Flask-ML API and Command Line Interface (CLI) support
• Ollama Inference: Uses Mistral 7B locally
• Chunked Inference: Handles long conversations by splitting into chunks of 30 messages
• Clean CSV Output with structured results

---

Getting Started

Prerequisites

Ensure you have the following installed:

• Python >=3.8
• Pip (latest version recommended)
• Ollama installed locally and running
• Mistral 7B model pulled via:

ollama pull mistral:7b-instruct

Clone the Repository

git clone https://github.com/mohanasrujana/Message-Analyzer.git
cd Message-Analyzer

Install Dependencies

pip install -r requirements.txt

---

Project Structure

📂 Message-Analyzer/
│
├── data/
|   ├── processed/
|   |   ├── combined_conversations.csv # Combined Dataset
|   ├── raw/
|   |   ├── true_negative_dataset/ #Prompt variables for true negative conversations
|   |   |   ├── augmented_true_negative_conversations.csv
|   |   |   ├── cities.txt
|   |   |   ├── conversation_topics.txt
|   |   |   ├── cross_validation_statistics.txt
|   |   |   ├── locations.txt
|   |   |   ├── mistral_validation_results.json
|   |   |   ├── participant_ages.txt
|   |   |   ├── participant_genders.txt
|   |   |   ├── participant_interests.txt
|   |   |   ├── participant_occupations.txt
|   |   |   ├── participant_personalities.txt
|   |   |   ├── time_settings.txt
|   |   |   ├── true_negative_conversations.csv
|   |   |   ├── true_negative_permutations.json
|   |   |   ├── true_negative_results.json
|   |   |   ├── true_negative_statistics.txt
|   |   ├── ambiguous_conversations.csv
|   |   ├── true_positives_conversations.csv
|   ├── combined_conversations_copy.csv
|   ├── predicted_result.csv
|   ├── error_log.txt
|
|
├── questions # (Not used in MVP; Plans to explore this option further down the line)
|   ├── investigative_questions.csv
|   ├── preset_questions.json
|
├── results/
|   ├── predicted_result.csv
|   ├── predicted_result_gemma_2b.csv
|   ├── raw_analysis.csv
|
├── scripts/
|   ├── api/
|   |   ├── cli.py
|   |   ├── server_info.md
|   |   ├── server.py
|   |   ├── server_gemma_2b.py
|   ├── jupyter_notebooks/
|   |   ├── Message_Analyser_Message_Generation.ipynb
|   |   ├── Message_Analyser_QuestionGeneration.ipynb
|   |   ├── Message_Analyzer_Ambiguous_Message_Generation.ipynb
|   |   ├── Message_Generation_True_Negative_.ipynb
|   |   ├── Mistral_Eval_True_Positives_.ipynb
|   |   ├── True_positives_ground_truth.ipynb
|   ├── export_to_onnx.py # (No longer required for MVP)
|   ├── extract_questions.py
|   ├── preprocessing.py
|   ├── run_mistral_inference.py
|   ├── run_onnx_inference.py
|
├── test/
|   ├── model_test.py # (No longer required for MVP)
|   ├── test_onnx_model.py # (No longer required for MVP)
|
├── .gitignore
├── requirements.txt  # Dependencies
├── README.md  # Documentation

---

Key Components

Model Export Script (export_to_onnx.py)

• Converts the modified Gemma:2b model to ONNX format.
• No longer used for MVP

ONNX Inference Script (run_onnx_inference.py)

• Loads the ONNX model and performs inference on a given conversation.
• This was part of the product building process; No longer used/required for MVP

Command Line Interface (CLI) (cli.py)

• Provides an easy-to-use CLI for message analyzisation.

Flask-ML API (server.py)

• Deploys the model as a web API.

---

Running Inference

CLI Usage

CLI help

python -m scripts.api.cli --help  

Predicting the conversations using cli:

Replace conversations_file with the input file of your conversations, results_dir with the directory in which you want your results, and "Actus Reus,Mens Rea" with the elements of crime you'd like to extract

python -m scripts.api.cli analyze --input_file [conversations_file] --output_file [results_dir] --elements_of_crime "Actus Reus,Mens Rea"

Here's the example command that worked for us:

python -m scripts.api.cli analyze --input_file data/combined_conversations_copy.csv --output_file results --elements_of_crime "Actus Reus,Mens Rea"

API Usage

Start the Flask-ML API server:

python scripts/api/server.py

Server usage (method 1)

Once running, send a POST request manually on the terminal:

curl -X POST "http://127.0.0.1:5000/analyze" \
     -H "Content-Type: multipart/form-data" \
     -F "input_file=@/path/to/conversations.csv" \
     -F "output_file=@/path/to/output/dir" \
     -F "elements_of_crime=Actus Reus,Mens Rea"

Server usage(method 2)

Use Rescue-Box-Desktop

• Install Rescue-Box from link
• Open Rescue-Box-Desktop and register the model by adding the server IP address and port number in which the server is running.
• Choose the model from list of available models under the MODELS tab.
• Checkout the Inspect page to learn more about using the model.
• Run the model.
• View the output in Jobs
• Click on view to view the details and results

---

Output Format

Your output will be a CSV with the following format:

conversation_id	chunk_id	message_number	speaker	message	crime_element
1	2	48	Marcus	Just... some guys moving stuff. Quickly.	Actus Reus

---

Future Enhancements

1. Add support for more models via toggle (e.g., Gemma ONNX, Mistral Ollama)

2. Improve hallucination filtering and accuracy

3. Improve the gemma model to give accurate results

Authors

1. Satya Srujana Pilli

2. Ashwini Ramesh Kumar

3. Shalom Jaison

---