final push

README.md

@@ -13,9 +13,100 @@ tags:
   - building-mcp-track-enterprise
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+
 
 
 link to social media : https://www.linkedin.com/feed/update/urn:li:activity:7400904066005966848/
 
-link to vidée : 
+link to vidéo : 
+
+# Project description
+
+This MCP server provides a multi-agent, tool-augmented reasoning system designed for clinical research, biomedical analysis, and scientific information retrieval. It integrates structured clinical trial querying, PubMed literature search, PDF parsing, RAG-based document retrieval, and web research into a single orchestrated environment.
+
+At its core is a Manager Agent (smolagents) that interprets user questions, plans multi-step reasoning, and coordinates two specialized agents:
+
+1. Clinical Agent
+
+Queries ClinicalTrials.gov with structured filters
+
+Retrieves PubMed articles
+
+Extracts text from scientific PDFs
+
+Produces structured tables, trial summaries, and evidence-based reports
+
+2. Online Information Agent
+
+Performs Wikipedia and DuckDuckGo searches
+
+Visits and extracts content from arbitrary webpages
+
+Supports general research, context building, and cross-verification
+
+## Summary
+
+This MCP server is a full-stack multi-agent research system with:
+
+Hierarchical LLM planning
+
+Dedicated scientific and clinical tools
+
+Real-time execution monitoring
+
+FAISS-based custom RAG infrastructure
+
+Integrated web search and document extraction
+
+A complete interactive UI for researchers or clinicians
+
+It is suitable for:
+
+Clinical evidence synthesis
+
+Scientific research workflows
+
+Medical question answering
+
+Literature reviews
+
+Automated extraction pipelines
+
+# Use Gradio Interface
+Use the specialized code agent from gradio interface [track agent reasonning]
+* Must Add NEBIUS_API_KEY secrets to the huggingface space in order to use the Agent
+* Add LANGFUSE secrets to track
+
+# Set-up MCP tools for a client (huggingchat)
+
+## 1. Server side : Connect to the space and click "Utiliser via API"
+![alt text](imgs/Step_1-2.jpg)
+## 2. Client side : Select "Manage MCP server" on huggingchat
+
+
+# --------------------------------------------------------
+
+
+## 3. Server side : Choose communication type (MCP streamable HTTP)
+![alt text](imgs/Step_3-4.jpg)
+## 4. Client side : Click on "Add Server"
+
+
+# ---------------------------------------------------------
+
+
+## 5. Server side : Copy the link for the client side (MCP streamable HTTP)
+![alt text](imgs/Step_5-6.jpg)
+## 6. Client side : Paste the link Click in "Server URL"
+
+
+# ----------------------------------------------------------
+
+
+
+
+# Chat with data
+ 
+## 7. Query LLM Without/With MCP tool changes the actions/output
+## 8. The MCP tools can be called multiple times (x6) in a single request
+![alt text](imgs/chat_with_data.jpg)

agent.py

@@ -19,19 +19,19 @@ from smolagents import (
 
 load_dotenv()
 
-from langfuse import get_client
-langfuse = get_client()
-if langfuse.auth_check():
-    print("Langfuse client is authenticated and ready!")
-else:
-    print("Authentication failed. Please check your credentials and host.")
+# from langfuse import get_client
+# langfuse = get_client()
+# if langfuse.auth_check():
+#     print("Langfuse client is authenticated and ready!")
+# else:
+#     print("Authentication failed. Please check your credentials and host.")
 
 
-from openinference.instrumentation.smolagents import SmolagentsInstrumentor
-SmolagentsInstrumentor().instrument()
+# from openinference.instrumentation.smolagents import SmolagentsInstrumentor
+# SmolagentsInstrumentor().instrument()
 
 model = LiteLLMModel(
-    model_id="openai/Qwen/Qwen3-Coder-480B-A35B-Instruct",
+    model_id="openai/nvidia/Llama-3_1-Nemotron-Ultra-253B-v1",
     api_key=os.environ.get("NEBIUS_API_KEY"),
     api_base="https://api.tokenfactory.nebius.com/v1/"
 )
@@ -113,6 +113,7 @@ clinical_agent = CodeAgent(
         "Gather general or recent information from online sources. "
         "Use Wikipedia for overviews, DuckDuckGo for recent data, and VisitWebpageTool for specific URLs. "
         "Return structured summaries with sources."
+        "Use the ClinicalTrialsSearchTool() for any question related to clinical trial"
     ),
     tools=[ClinicalTrialsSearchTool()],
     additional_authorized_imports=["time", "numpy", "pandas"],
@@ -150,10 +151,11 @@ manager_agent = CodeAgent(
     "Most important task is to provide a complete answer to user questions based on clinical trial data and online information. "
     "Orchestrate workflow between clinical and online agents. "
     "Validate outputs, resolve conflicts, and ensure the final answer is complete and accurate."
+    "rimarily use the managed agent clinical_agent for question related to clinical trials"
     ),
-    tools=[FinalAnswerTool()],
+    tools=[FinalAnswerTool(),ClinicalTrialsSearchTool()],
     model=model,
-    managed_agents=[clinical_agent,search_online_info],
+    # managed_agents=[clinical_agent,search_online_info],
     # executor_type="modal",
     provide_run_summary=True,
     additional_authorized_imports=["time", "numpy", "pandas"],

app.py

@@ -83,7 +83,7 @@ def answer_question(question, history):
         question_with_history = "Conversation history:\n" + str(history) + "\n\nNew user question:\n " + question
         for st in safe_agent.run(question_with_history,stream=True,return_full_result=True):
             if isinstance(st, smolagents.memory.PlanningStep):
-                plan = "# Plan" + st.plan.split("## 2. Plan")[-1]
+                plan = 20*"# " + "\n# Planning of manager agent" + st.plan.split("## 2. Plan")[-1]
                 for m in plan.split("\n"):
                     thoughts += "\n" + m
                     yield thoughts, final_answer, history
@@ -103,13 +103,20 @@ def answer_question(question, history):
                     yield thoughts, final_answer, history
 
             elif isinstance(st,  smolagents.memory.ActionStep):
-                if st.observations:
-                    for m in st.observations.split("\n")[1:]:
-                        thoughts += m
+        
+                for chatmessage in st.model_input_messages:
+                    if chatmessage.role == "assistant":
+                        managed_agent_plan = chatmessage.content[0]['text'].split("2. Plan")[-1]
+                        thoughts += "Managed agent plan:\n"
+                        for l in managed_agent_plan.split("\n"):
+                            thoughts += l
+                        thoughts += "\n\n--> Code action from managed agent \n" + st.code_action +"\n\n"
                         yield thoughts, final_answer, history
-
-                thoughts += "\n\n\n********** End fo Step " + str(st.step_number) + " : *********\n" + str(st.token_usage) + "\nStep duration" + str(st.timing) + "\n\n"
+                thoughts += "\n********** End fo Step " + str(st.step_number) + " : *********\n" + str(st.token_usage) + "\nStep duration" + str(st.timing) + "\n\n"
                 yield thoughts, final_answer, history
+
+
+
             elif isinstance(st, smolagents.memory.FinalAnswerStep):
                 final_answer = st.output
                 history.append({"question": question, "answer": final_answer})
@@ -123,13 +130,6 @@ def answer_question(question, history):
         return "Request cancelled","Submit new request", ""
 
 
-
-# def create_rag_files(refs :list[str], VECTOR_DB_PATH:str)-> str:
-#     from tool_create_FAISS_vector import create_vector_store_from_list_of_doi
-
-#     FAISS_VECTOR_PATH = create_vector_store_from_list_of_doi(refs,VECTOR_DB_PATH)
-#     return FAISS_VECTOR_PATH
-
 def tool_clinical_trial(query_cond:str=None, query_term:str=None,query_lead:str=None,max_results: str="5") -> str:
     """
     Search Clinical Trials database for trials with 4 arguments.
@@ -196,6 +196,7 @@ def create_rag(refs :str, VECTOR_DB_PATH:str)-> str:
     """
     from tool_create_FAISS_vector import create_vector_store_from_list_of_doi
     FAISS_VECTOR_PATH = create_vector_store_from_list_of_doi(refs,VECTOR_DB_PATH)
+
     return FAISS_VECTOR_PATH
 
 
@@ -207,11 +208,11 @@ def use_rag(query: str, store_name: str, top_k: int = 5) -> str:
         store_name (str): The path to the FAISS vector store to query.
         top_k (int): The number of top-k most relevant context documents to retrieve (default: 5).
     Returns:
-        str: A JSON string containing the retrieved context, including the content and source (DOI).
+        str: A TOON formated string containing the retrieved contexts, including the contents, the source and the scores.
     """
     from tool_query_FAISS_vector import query_vector_store
     context_as_dict = query_vector_store(query, store_name, top_k)
-    return json.dumps(context_as_dict, indent=2)
+    return str(context_as_dict)
 
 
 def describe_figure(image : Image.Image) -> str:
@@ -230,14 +231,6 @@ def describe_figure(image : Image.Image) -> str:
 # Create neat interface - Question Analyzer as a Blocks component
 with gr.Blocks() as interface2:
     gr.Markdown("# Question Analyzer")
-    # gr.Markdown("""Enter a question to analyze. Examples:
-    # - Find the name of the sponsor that did the most studies on Alzheimer's disease in the last 10 years.
-    # - Provide a summary of recent clinical trials on diabetes and list 3 relevant research articles from PubMed.
-    # - What are the scientific paper linked to the clinical study referenced as NCT04516746?
-    # - How many clinical studies on cancer were completed in the last 5 years?
-    # - Find recent phase 3 trials for lung cancer sponsored by Pfizer
-    # """)
-    
     with gr.Row():
         with gr.Column():
             question_input = gr.Textbox(
@@ -246,9 +239,9 @@ with gr.Blocks() as interface2:
                 lines=3,
             )
             gr.Examples(["What is the weather in LA?",
-                         "What are the pmid of the study NCT04516746?",
-                         "How many studies on cancer were completed in the last 5 years?",
-                         "Find phase 3 trials for lung cancer sponsored by Pfizer"],question_input)
+                         "What are the 5 most recent clinical study sponsored by Merck?",
+                         "How many trials were completed in 2025 by AbbVie?",
+                         "What are the pmids associated with the study NCT04516746?",],question_input)
             with gr.Row():
                 submit_btn = gr.Button("Submit", variant="primary")
                 stop_btn   = gr.Button("Stop", variant="secondary") 
@@ -334,7 +327,8 @@ with gr.Blocks() as interface3:
                 inputs=[ref_input, vector_name_input]
             )
             path_output = gr.Textbox(
-                label="Path of the vector store"
+                label="Path of the vector store",
+                lines=4
             )
             submit_btn = gr.Button("Create Vector Store")
             submit_btn.click(
@@ -353,7 +347,7 @@ demo = gr.TabbedInterface(
             fn=use_rag, 
             inputs=[gr.Textbox(label="Question that needs context to answer", placeholder="What is the dose of medicine to gove an infant under type2 diabetes"), 
                     gr.Textbox(label="Name of the vector store to use", placeholder="Diabetes, Sickel_cell_anemia, Prostate_cancer, ..")], 
-            outputs=gr.Textbox(label="Answer with Rag",placeholder="Your answer will be provided here"),
+            outputs=gr.Textbox(label="Answer with Rag",lines=8, placeholder="Your answer will be provided here"),
             api_name="use_vector_store_to_create_context"),
          gr.Interface(
             fn=tool_clinical_trial, 

test_nebius_mode.ipynb

@@ -0,0 +1,105 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "id": "371b903a",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Available models:\n",
+      "- meta-llama/Meta-Llama-3.1-8B-Instruct-fast\n",
+      "- meta-llama/Meta-Llama-3.1-8B-Instruct\n",
+      "- meta-llama/Llama-Guard-3-8B\n",
+      "- nvidia/Llama-3_1-Nemotron-Ultra-253B-v1\n",
+      "- nvidia/Nemotron-Nano-V2-12b\n",
+      "- google/gemma-2-2b-it\n",
+      "- google/gemma-2-9b-it-fast\n",
+      "- Qwen/Qwen2.5-Coder-7B-fast\n",
+      "- BAAI/bge-en-icl\n",
+      "- BAAI/bge-multilingual-gemma2\n",
+      "- intfloat/e5-mistral-7b-instruct\n",
+      "- meta-llama/Llama-3.3-70B-Instruct\n",
+      "- meta-llama/Llama-3.3-70B-Instruct-fast\n",
+      "- deepseek-ai/DeepSeek-R1-0528\n",
+      "- Qwen/Qwen3-235B-A22B-Instruct-2507\n",
+      "- Qwen/Qwen3-235B-A22B-Thinking-2507\n",
+      "- Qwen/Qwen3-32B\n",
+      "- Qwen/Qwen3-32B-fast\n",
+      "- google/gemma-3-27b-it\n",
+      "- google/gemma-3-27b-it-fast\n",
+      "- Qwen/Qwen2.5-VL-72B-Instruct\n",
+      "- Qwen/Qwen3-Embedding-8B\n",
+      "- deepseek-ai/DeepSeek-V3-0324\n",
+      "- zai-org/GLM-4.5\n",
+      "- zai-org/GLM-4.5-Air\n",
+      "- openai/gpt-oss-120b\n",
+      "- openai/gpt-oss-20b\n",
+      "- deepseek-ai/DeepSeek-R1-0528-fast\n",
+      "- deepseek-ai/DeepSeek-V3-0324-fast\n",
+      "- moonshotai/Kimi-K2-Instruct\n",
+      "- Qwen/Qwen3-30B-A3B-Thinking-2507\n",
+      "- Qwen/Qwen3-30B-A3B-Instruct-2507\n",
+      "- Qwen/Qwen3-Coder-30B-A3B-Instruct\n",
+      "- Qwen/Qwen3-Coder-480B-A35B-Instruct\n",
+      "- NousResearch/Hermes-4-70B\n",
+      "- NousResearch/Hermes-4-405B\n",
+      "- PrimeIntellect/INTELLECT-3\n",
+      "- moonshotai/Kimi-K2-Thinking\n",
+      "- black-forest-labs/flux-dev\n",
+      "- black-forest-labs/flux-schnell\n"
+     ]
+    }
+   ],
+   "source": [
+    "import requests\n",
+    "import os\n",
+    "\n",
+    "api_key = os.environ.get(\"NEBIUS_API_KEY\")\n",
+    "headers = {\n",
+    "    \"Authorization\": f\"Bearer {api_key}\",\n",
+    "    \"Content-Type\": \"application/json\"\n",
+    "}\n",
+    "\n",
+    "# Try the models endpoint\n",
+    "response = requests.get(\n",
+    "    \"https://api.tokenfactory.nebius.com/v1/models\",\n",
+    "    headers=headers\n",
+    ")\n",
+    "\n",
+    "if response.status_code == 200:\n",
+    "    models = response.json()\n",
+    "    print(\"Available models:\")\n",
+    "    for model in models.get('data', []):\n",
+    "        print(f\"- {model.get('id')}\")\n",
+    "else:\n",
+    "    print(f\"Error: {response.status_code}\")\n",
+    "    print(response.text)"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "mcp-birthday",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.19"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

tool_create_FAISS_vector.py

@@ -369,9 +369,9 @@ from transformers import AutoTokenizer
 from langchain_text_splitters import RecursiveCharacterTextSplitter
 from tqdm import tqdm
 
-def create_vector_store_from_list_of_doi(refs :str, VECTOR_DB_PATH:str) -> str:
+def create_vector_store_from_list_of_doi(refs :str, VECTOR_DB_NAME:str) -> str:
 
-    VECTOR_DB_PATH = "./tmp/vector_stores/" + VECTOR_DB_PATH
+    VECTOR_DB_PATH = "./tmp/vector_stores/" + VECTOR_DB_NAME
     
     from langchain_community.vectorstores import FAISS
 
@@ -434,10 +434,25 @@ def create_vector_store_from_list_of_doi(refs :str, VECTOR_DB_PATH:str) -> str:
             print("merge vector store")
             KNOWLEDGE_VECTOR_DATABASE.merge_from(NEW_KNOWLEDGE_VECTOR_DATABASE)
             KNOWLEDGE_VECTOR_DATABASE.save_local(VECTOR_DB_PATH)
+            vector_type={"name":VECTOR_DB_NAME,
+                        "num_vectors" : KNOWLEDGE_VECTOR_DATABASE.index.ntotal,
+                        "vector_dim" : KNOWLEDGE_VECTOR_DATABASE.index.d,
+                        "distance_strategy" : KNOWLEDGE_VECTOR_DATABASE.distance_strategy}
         else:
             NEW_KNOWLEDGE_VECTOR_DATABASE.save_local(VECTOR_DB_PATH)
 
-        return VECTOR_DB_PATH
+            vector_type={"name":VECTOR_DB_NAME,
+                         "num_vectors" : NEW_KNOWLEDGE_VECTOR_DATABASE.index.ntotal,
+                        "vector_dim" : NEW_KNOWLEDGE_VECTOR_DATABASE.index.d,
+                        "distance_strategy" : NEW_KNOWLEDGE_VECTOR_DATABASE.distance_strategy}
+        return str(vector_type)
     
     else:
-        return f"all the data already in vector store {VECTOR_DB_PATH}"
+        if KNOWLEDGE_VECTOR_DATABASE:
+            vector_type={"name":VECTOR_DB_NAME,
+                            "num_vectors" : KNOWLEDGE_VECTOR_DATABASE.index.ntotal,
+                            "vector_dim" : KNOWLEDGE_VECTOR_DATABASE.index.d,
+                            "distance_strategy" : KNOWLEDGE_VECTOR_DATABASE.distance_strategy}
+            return str(vector_type)
+        else:
+            return f"could not extract new ref for this vector store"

tool_query_FAISS_vector.py

@@ -44,14 +44,17 @@ def query_vector_store(query: str, store_name: str, top_k: int = 5) -> dict:
         embedding_model, 
         allow_dangerous_deserialization=True  
         )
-
-    results = vector_store.similarity_search(query, top_k)
-    
-    context = [r.page_content for r in results]
-    sources = [r.metadata["source"] for r in results]
     
-    return {
+    results = vector_store.similarity_search_with_score(query, top_k)
+
+    context = [r[0].page_content for r in results][::-1]
+    sources = [r[0].metadata["source"] for r in results][::-1]
+    scores = [r[1] for r in results][::-1]
+
+    dict_ = {
         "context": context,
         "sources": sources,
-        "store_name": store_name
-    }
+        "scores": scores,
+        "store_name": "s"
+    }
+    return dict_