Merge pull request #56 from boettiger-lab/add_llms

update llm options and adding logging option

Author: cassiebuhler

app/app.py

@@ -129,9 +129,10 @@
 #############
 
 
+
 chatbot_container = st.container()
 with chatbot_container:
-    llm_left_col, llm_right_col = st.columns([5,1], vertical_alignment = "bottom")
+    llm_left_col, llm_right_col = st.columns([2.5,1], vertical_alignment = "bottom")
     with llm_left_col:
 
         with st.popover("💬 Example Queries"):
@@ -157,7 +158,6 @@
 
             st.info('If the map appears blank, queried data may be too small to see at the default zoom level. Check the table below the map, as query results will also be displayed there.', icon="ℹ️")
 
-
     with llm_right_col:
         llm_choice = st.selectbox("Select LLM:", llm_options, key = "llm", help = "Select which model to use.")   
         llm = llm_options[llm_choice]
@@ -206,7 +206,7 @@ def run_sql(query,color_choice):
     explanation =output.explanation
     if not sql_query: # if the chatbot can't generate a SQL query.
         st.success(explanation)
-        return pd.DataFrame({'id' : []})
+        return pd.DataFrame({'id' : []}), [], []
 
     result = ca.sql(sql_query).execute()
     if result.empty :
@@ -215,9 +215,9 @@ def run_sql(query,color_choice):
         st.caption("SQL Query:")
         st.code(sql_query,language = "sql") 
         if 'geom' in result.columns:
-            return result.drop('geom',axis = 1)
+            return result.drop('geom',axis = 1),  sql_query, explanation
         else: 
-            return result
+            return result, sql_query, explanation
 
     elif ("id" and "geom" in result.columns): 
         style = get_pmtiles_style_llm(style_options[color_choice], result["id"].tolist())
@@ -235,7 +235,7 @@ def run_sql(query,color_choice):
         st.caption("SQL Query:")
         st.code(sql_query,language = "sql") 
 
-    return result
+    return result, sql_query, explanation
 
 
 
@@ -266,15 +266,22 @@ def run_sql(query,color_choice):
         example_query = "👋 Input query here"
         prompt = st.chat_input(example_query, key="chain", max_chars=300)
 
+with chatbot_container:      
+        _,log_query_col, _ = st.columns([.001, 5,1], vertical_alignment = "top")
+        with log_query_col:
+            log_queries = st.checkbox("Save query", value = True, help = "Saving your queries helps improve this tool and guide conservation efforts. Your data is stored in a private location. For more details, see 'Why save your queries?' at the bottom of this page.")
 
+    
 with st.container():
     if prompt: 
         st.chat_message("user").write(prompt)
         try:
             with st.chat_message("assistant"):
                 with st.spinner("Invoking query..."):
 
-                    out = run_sql(prompt,color_choice)
+                    out, sql_query, llm_explanation = run_sql(prompt,color_choice)
+                    minio_logger(log_queries, prompt, sql_query, llm_explanation, llm_choice, 'query_log_prototype.csv', "shared-ca30x30-app")
+                    
                     if ("id" in out.columns) and (not out.empty):
                         ids = out['id'].tolist()
                         cols = out.columns.tolist()

app/footer.md

@@ -16,3 +16,17 @@ Data: https://huggingface.co/datasets/boettiger-lab/ca-30x30
 - CDC 2022 Social Vulnerability Index by US Census Tract. Archived description: https://web.archive.org/web/20250126095916/https://www.atsdr.cdc.gov/place-health/php/svi/index.html. Data: https://source.coop/repositories/cboettig/social-vulnerability/description. License: Public Domain
 
 - Fire and Prescribed Fire by CAL FIRE (2023), reprocessed to PMTiles on https://beta.source.coop/cboettig/fire/. License: Public Domain
+
+
+#### LLMs
+This app can use a selection of open-weights language models hosted on the National Research Platform (https://nrp.ai/documentation/userdocs/ai/llm-managed/), and Open Router (https://openrouter.ai/models).
+
+---
+
+### Why save your queries?
+Conservation researchers and practitioners are interested in **learning what matters most to our community**. 
+
+By saving your anonymous queries, we can identify which topics and areas are drawing the most attention, helping us improve future tools and data products to **better support conservation efforts**. We also save the LLM’s response to each query to monitor its accuracy and ensure the system is working as intended. 
+
+You can opt out at any time by disabling “Save query”.
+

app/utils.py

@@ -11,6 +11,8 @@
 from typing import Optional
 from functools import reduce
 from itertools import chain
+import minio
+import datetime
 
 from variables import *
 
@@ -407,3 +409,28 @@ def create_bar_chart(df, x, y, title, color=None, stacked=False, colors=None):
     )
 
     return final_chart
+
+minio_key = os.getenv("MINIO_KEY")
+if minio_key is None:
+    minio_key = st.secrets["MINIO_KEY"]
+
+minio_secret = os.getenv("MINIO_SECRET")
+if minio_secret is None:
+    minio_secret = st.secrets["MINIO_SECRET"]
+
+def minio_logger(consent, query, sql_query, llm_explanation, llm_choice, filename, bucket,
+                 key=minio_key, secret=minio_secret,
+                 endpoint="minio.carlboettiger.info"):
+    mc = minio.Minio(endpoint, key, secret)
+    mc.fget_object(bucket, filename, filename)
+    log = pd.read_csv(filename)
+    timestamp = datetime.datetime.now(datetime.timezone.utc).strftime("%Y-%m-%d %H:%M:%S")
+    if consent:
+        df = pd.DataFrame({"timestamp": [timestamp], "user_query": [query], "llm_sql": [sql_query], "llm_explanation": [llm_explanation], "llm_choice":[llm_choice]})
+
+    # if user opted out, do not store query
+    else:  
+        df = pd.DataFrame({"timestamp": [timestamp], "user_query": ['USER OPTED OUT'], "llm_sql": [''], "llm_explanation": [''], "llm_choice":['']})
+    
+    pd.concat([log,df]).to_csv(filename, index=False, header=True)
+    mc.fput_object(bucket, filename, filename, content_type="text/csv")

app/variables.py

@@ -293,13 +293,53 @@
 
 from langchain_openai import ChatOpenAI
 import streamlit as st
-# from langchain_openai.chat_models.base import BaseChatOpenAI
+from langchain_openai.chat_models.base import BaseChatOpenAI
 
-llm_options = {
-    # "llama-3.3-quantized": ChatOpenAI(model = "cirrus", api_key=st.secrets['CIRRUS_LLM_API_KEY'], base_url = "https://llm.cirrus.carlboettiger.info/v1",  temperature=0),
-    "llama3.3": ChatOpenAI(model = "llama3-sdsc", api_key=st.secrets['NRP_API_KEY'], base_url = "https://llm.nrp-nautilus.io/",  temperature=0),
-    "gemma3": ChatOpenAI(model = "gemma3", api_key=st.secrets['NRP_API_KEY'], base_url = "https://llm.nrp-nautilus.io/",  temperature=0),
-    # "DeepSeek-R1-Distill-Qwen-32B": BaseChatOpenAI(model = "DeepSeek-R1-Distill-Qwen-32B", api_key=st.secrets['NRP_API_KEY'], base_url = "https://llm.nrp-nautilus.io/",  temperature=0),
-    "watt": ChatOpenAI(model = "watt", api_key=st.secrets['NRP_API_KEY'], base_url = "https://llm.nrp-nautilus.io/",  temperature=0),
-    # "phi3": ChatOpenAI(model = "phi3", api_key=st.secrets['NRP_API_KEY'], base_url = "https://llm.nrp-nautilus.io/",  temperature=0),
+## dockerized streamlit app wants to read from os.getenv(), otherwise use st.secrets
+import os
+api_key = os.getenv("NRP_API_KEY")
+if api_key is None:
+    api_key = st.secrets["NRP_API_KEY"]
+
+openrouter_api = os.getenv("OPENROUTER_API_KEY")
+if openrouter_api is None:
+    openrouter_api = st.secrets["OPENROUTER_API_KEY"]
+
+openrouter_endpoint="https://openrouter.ai/api/v1"
+nrp_endpoint="https://ellm.nrp-nautilus.io/v1"
+
+# don't use a provider that collects data
+data_policy = {
+    "provider": {
+        "data_collection": "deny"
+    }
 }
+
+llm_options = {
+    "gemma-3-27b-it": ChatOpenAI(
+        model="gemma3",
+        api_key=api_key,
+        base_url=nrp_endpoint,
+        temperature=0
+    ),
+    "gpt-oss-120b": ChatOpenAI(
+        model="gpt-oss",
+        api_key=api_key,
+        base_url=nrp_endpoint,
+        temperature=0
+    ),
+    "trinity-mini": ChatOpenAI(
+        model="arcee-ai/trinity-mini:free",
+        api_key=openrouter_api,
+        base_url=openrouter_endpoint,
+        temperature=0,
+        extra_body=data_policy
+    ),
+    "nemotron-nano-9b-v2": ChatOpenAI(
+        model="nvidia/nemotron-nano-9b-v2:free",
+        api_key=openrouter_api,
+        base_url=openrouter_endpoint,
+        temperature=0,
+        extra_body=data_policy
+    ),
+}

requirements.txt

@@ -13,7 +13,5 @@ geoarrow-pandas==0.1.1
 leafmap==0.38.12 
 SQLAlchemy==2.0.35
 duckdb_engine==0.14.2
-langchain==0.2.17
-langchain-community==0.2.19
-langchain-core==0.2.43
-langchain-openai==0.1.25
+langchain-core==1.1.0
+langchain-openai==1.1.0