main.py

from flask import Flask, request, jsonify
import ollama

app = Flask(__name__)


@app.route('/generate', methods=['POST'])
def generate_response():
    try:
        # Get input text from the request
        input_text = request.json.get('input_text')
        # Ensure the user message contains content
        if input_text:
            # Send user message through the chat stream
            # chat_stream.send(messages = {'role': 'user', 'content': input_text})
            chat_stream = ollama.chat(
    model="mistral",
    messages=[{'role': 'user', 'content': input_text}], 
    stream=True
)

            print('text sent to model ')
            # Collect response chunks from the stream
            response_chunks = []
            for chunk in chat_stream:
                response_chunks.append(chunk['message']['content'])

                # Check for Mistral's "end_of_response" signal
                if chunk['message']['role'] == 'system' and chunk['message'].get('end_of_response'):
                    break

            # Complete response
            response = ''.join(response_chunks)
            print('response ' , response)
            return jsonify({"response": response})
        else:
            return jsonify({"error": "User message must contain content."})

    except Exception as e:
        return jsonify({"error": str(e)})

if __name__ == '__main__':
    app.run(host='0.0.0.0', port=8000)