Create utils.py

utils.py

@@ -0,0 +1,55 @@
+# utils.py
+import pandas as pd
+from langchain.llms import HuggingFaceHub
+from langchain.agents import create_pandas_dataframe_agent
+from typing import Tuple
+
+def query_agent_from_csv(file_bytes, query: str, hf_token: str, repo_id: str = "mistralai/mistral-7b-instruct") -> Tuple[str, str]:
+    """
+    Reads a CSV from file-like bytes, builds a LangChain pandas-agent with HuggingFaceHub LLM,
+    runs the user query and returns (answer, debug_info).
+    - file_bytes: bytes of the uploaded CSV file (Streamlit provides)
+    - query: user's natural language question
+    - hf_token: huggingface token (string)
+    - repo_id: huggingface repo id for the model to use (e.g. 'mistralai/mistral-7b-instruct')
+    Returns: tuple (answer_text, debug_text)
+    """
+    try:
+        # Read CSV — attempt common encodings and fallback
+        try:
+            df = pd.read_csv(file_bytes)
+        except Exception:
+            # try with latin1 encoding as fallback
+            file_bytes.seek(0)
+            df = pd.read_csv(file_bytes, encoding="latin1")
+    except Exception as e:
+        return f"Error reading CSV: {e}", ""
+
+    # small safety: if dataset is extremely wide, limit columns
+    MAX_COLS = 200
+    if df.shape[1] > MAX_COLS:
+        df = df.iloc[:, :MAX_COLS]
+
+    # Build the LLM wrapper for Hugging Face Hub
+    try:
+        llm = HuggingFaceHub(
+            repo_id=repo_id,
+            huggingfacehub_api_token=hf_token,
+            model_kwargs={"temperature": 0.0, "max_new_tokens": 512},
+        )
+    except Exception as e:
+        return "", f"Error creating HuggingFaceHub LLM: {e}"
+
+    # Create pandas agent
+    try:
+        agent = create_pandas_dataframe_agent(llm, df, verbose=False)
+    except Exception as e:
+        return "", f"Error creating LangChain pandas agent: {e}"
+
+    # Run query (wrap in try/except to capture agent errors)
+    try:
+        answer = agent.run(query)
+    except Exception as e:
+        return "", f"Agent runtime error: {e}"
+
+    return answer, ""