ecologits codecarbon

app.py

@@ -40,6 +40,15 @@ def main():
         tracker = EmissionsTracker()
         tracker.start()
         st.session_state["emission"] = tracker
+    
+    if "partial_emissions" not in st.session_state:
+        st.session_state["partial_emissions"] = {
+            "Scrapping": {"cc": 0, "el": 0}, #cc for codecarbon , el for ecologits
+            "extraction_pp": {"cc": 0, "el": 0},
+            "cartographie": {"cc": 0, "el": 0},
+            "chatbot": {"cc": 0, "el": 0},
+            "download_rapport": {"cc": 0, "el": 0},
+        }
  
     st.sidebar.title("OPEN DATA & IA au service de la RSE")
     section_principale = st.sidebar.radio(
@@ -137,17 +146,29 @@ def main():
     """,
     unsafe_allow_html=True
     )
-    st.sidebar.markdown(f"<div style='text-align: center; margin-bottom: 10px;'><b>{st.session_state['emission'].stop()*1000:.3f}</b> g eq. CO2</div>", unsafe_allow_html=True)
-    st.sidebar.markdown("(avec les outils code carbon, Ecologits)")
+    
+    serveur_emission = st.session_state['emission'].stop()
+    emission_api = sum([value["el"] for value in st.session_state["partial_emissions"].values()])
+
+    total_emission = serveur_emission + emission_api
+
+    pourcentage_api = emission_api / total_emission
+    pourcentage_serveur = serveur_emission / total_emission
+
+    st.sidebar.markdown(f"<div style='text-align: center; margin-bottom: 10px;'><b>{total_emission*1000:.3f}</b> g eq. CO2</div>", unsafe_allow_html=True)
+    st.sidebar.markdown("Dont :")
+    st.sidebar.markdown(f"- Empreinte serveur (via CodeCarbon) : **{serveur_emission*1000:.3f}** g eq. CO2 ({pourcentage_serveur:.2%})")
+    st.sidebar.markdown(f"- Empreinte serveur (via EcoLogits) : **{emission_api*1000:.3f}** g eq. CO2 ({pourcentage_api:.2%})")
+    # st.sidebar.markdown("(avec l'outil CodeCarbon)")
     c1,c2,c3 = st.sidebar.columns([1,1,1])
 
     c2.write("---")
 
-    st.sidebar.markdown("**Votre équivalence:**")
+    st.sidebar.markdown("**Votre équivalence**")
     col1,col2,col3 = st.sidebar.columns([1,1,1])
-    display_comparaison(col1,st.session_state['emission'].stop(),dict_comparaison_1kgCO2["eau en litre"][0]*1000,dict_comparaison_1kgCO2["eau en litre"][1],"ml")
-    display_comparaison(col2,st.session_state['emission'].stop(),dict_comparaison_1kgCO2["tgv en km"][0],dict_comparaison_1kgCO2["tgv en km"][1],"km")
-    display_comparaison(col3,st.session_state['emission'].stop(),dict_comparaison_1kgCO2["voiture en km"][0]*1000,dict_comparaison_1kgCO2["voiture en km"][1],"m")
+    display_comparaison(col1,total_emission,dict_comparaison_1kgCO2["eau en litre"][0]*1000,dict_comparaison_1kgCO2["eau en litre"][1],"ml")
+    display_comparaison(col2,total_emission,dict_comparaison_1kgCO2["tgv en km"][0],dict_comparaison_1kgCO2["tgv en km"][1],"km")
+    display_comparaison(col3,total_emission,dict_comparaison_1kgCO2["voiture en km"][0]*1000,dict_comparaison_1kgCO2["voiture en km"][1],"m")
     st.sidebar.markdown("\n")
     st.sidebar.markdown(
     f"""

chat_with_pps.py

@@ -13,6 +13,9 @@ import random
 import pandas as pd
 from codecarbon import EmissionsTracker
 
+from ecologits.tracers.utils import compute_llm_impacts
+import time
+import itertools
 
 
 
@@ -105,6 +108,58 @@ def get_response(user_query, chat_history, context,llm=None,history_limit=5,stre
         "user_question": user_query,
     })
 
+def get_response_with_impact(user_query, chat_history, context,llm=None,history_limit=5,stream=True):
+    model_vs_provider = {
+        "Mistral (FR)": ["mistral-large-latest","mistralai"],
+        "GPT-4o": ["gpt-4o","openai"]
+    }
+
+    start = time.perf_counter()
+    response_generator = get_response(user_query, chat_history, context,llm,history_limit,stream)
+    request_latency_invoke = time.perf_counter() - start
+    wrapped_response_generator, token_count_generator = itertools.tee(response_generator)
+
+    token_count = 0
+
+    
+    # Stream the response while counting tokens
+    if stream:
+        try:
+            final_response = st.write_stream(wrapped_response_generator)
+            
+        finally:
+            token_count = len(list(token_count_generator))
+    else:
+        final_response = response_generator
+        token_count = len(list(token_count_generator))
+    
+    request_latency_stream = time.perf_counter() - start
+
+    nbre_out_tokens = token_count
+    model_name = model_vs_provider[st.session_state.model][0]
+    model_provider = model_vs_provider[st.session_state.model][1]
+
+    if not stream:
+        request_latency = request_latency_invoke
+    else:
+        request_latency = request_latency_stream
+
+    impact = compute_llm_impacts(
+        provider=model_provider,
+        model_name=model_name,
+        output_token_count=nbre_out_tokens,
+        request_latency=request_latency,
+    )
+    
+    print(f"Request latency: {request_latency:.3f} s")
+    print(f"Output token count: {nbre_out_tokens}")
+    print(f"Impact: {impact.gwp.value} {impact.gwp.unit}")
+
+    st.session_state["partial_emissions"]["chatbot"]["el"] += impact.gwp.value
+
+    return final_response
+
+
 def display_chart():
     if "pp_grouped" not in st.session_state or st.session_state['pp_grouped'] is None or len(st.session_state['pp_grouped']) == 0:
         st.warning("Aucune partie prenante n'a été définie")
@@ -315,10 +370,10 @@ def display_chat():
                 st.markdown(f"**{st.session_state.model}**")
                 if st.session_state.model not in st.session_state.used_models:
                     st.session_state.used_models.append(st.session_state.model)
-                response = st.write_stream(get_response(last_message.content, st.session_state.chat_history,format_context(st.session_state['pp_grouped'],st.session_state['Nom de la marque']),st.session_state.model))
+                response = get_response_with_impact(last_message.content, st.session_state.chat_history,format_context(st.session_state['pp_grouped'],st.session_state['Nom de la marque']),st.session_state.model)
                 st.session_state.chat_history.append(AIMessage(content=response))
                 with st.spinner("Proposition de prompts..."):
-                    propositions_prompts = get_response(generated_prompt_question, st.session_state.chat_history,format_context(st.session_state['pp_grouped'],st.session_state['Nom de la marque']),st.session_state.model,history_limit=1,stream=False)
+                    propositions_prompts = get_response_with_impact(generated_prompt_question, st.session_state.chat_history,format_context(st.session_state['pp_grouped'],st.session_state['Nom de la marque']),st.session_state.model,history_limit=1,stream=False)
                 extract_format_prompts_from_response(propositions_prompts)
 
 
@@ -343,10 +398,13 @@ def display_chat():
                 format_pp_add_viz(partie_prenante)
                 disp_carto_in_chat()
             else:
-                response = st.write_stream(get_response(user_query, st.session_state.chat_history,format_context(st.session_state['pp_grouped'],st.session_state['Nom de la marque']),st.session_state.model))
+                #response = st.write_stream(get_response(user_query, st.session_state.chat_history,format_context(st.session_state['pp_grouped'],st.session_state['Nom de la marque']),st.session_state.model))
+                response = get_response_with_impact(user_query, st.session_state.chat_history,format_context(st.session_state['pp_grouped'],st.session_state['Nom de la marque']),st.session_state.model)
+
                 st.session_state.chat_history.append(AIMessage(content=response))
+
                 with st.spinner("Proposition de prompts..."):
-                    propositions_prompts = get_response(generated_prompt_question, st.session_state.chat_history,format_context(st.session_state['pp_grouped'],st.session_state['Nom de la marque']),st.session_state.model,history_limit=1,stream=False)
+                    propositions_prompts = get_response_with_impact(generated_prompt_question, st.session_state.chat_history,format_context(st.session_state['pp_grouped'],st.session_state['Nom de la marque']),st.session_state.model,history_limit=1,stream=False)
                 extract_format_prompts_from_response(propositions_prompts)
                 
             

partie_prenante_carte.py

@@ -3,6 +3,7 @@ import pandas as pd
 import numpy as np
 import re
 import random
+import time
 
 import streamlit as st
 from dotenv import load_dotenv
@@ -17,7 +18,7 @@ from langchain_core.prompts.prompt import PromptTemplate
 import os
 from high_chart import test_chart
 from chat_with_pps import get_response
-
+from ecologits.tracers.utils import compute_llm_impacts
 from codecarbon import EmissionsTracker
 
 load_dotenv()
@@ -184,14 +185,28 @@ def extract_pp(docs,input_variables):
 
     question = fill_promptQ_template(input_variables, template_extraction_PP)
 
+    start = time.perf_counter()
     response = chain.invoke(question)
 
+    response_latency = time.perf_counter() - start
     # version plus poussée a considérer
     # each item in the list is a list with the name of the brand and the similarity percentage
     # partie_prenante = text_to_list(response.content)
     if "ne sais pas" in response.content:
         return "444"
 
+    #calculate impact
+    nbre_out_tokens = response.response_metadata["token_usage"]["completion_tokens"]
+    provider = "openai"
+    model = "gpt-4o"
+    impact = compute_llm_impacts(
+        provider=provider,
+        model_name=model,
+        output_token_count=nbre_out_tokens,
+        request_latency=response_latency,
+    )
+
+    st.session_state["partial_emissions"]["extraction_pp"]["el"] += impact.gwp.value
     #version simple
     partie_prenante = response.content.replace("- ","").split('\n')
     partie_prenante = [item.strip() for item in partie_prenante]