Upload app.py

app.py

@@ -3,8 +3,13 @@ from llama_cpp import Llama
 from huggingface_hub import hf_hub_download
 import os
 import pickle
+import logging
 from langchain_huggingface import HuggingFaceEmbeddings
 
+# ====================== LOGGING ======================
+logging.basicConfig(level=logging.INFO, format="%(levelname)s | %(message)s")
+logger = logging.getLogger(__name__)
+
 # ====================== CONFIG ======================
 repo_id = "robertolofaro/articles-model"
 
@@ -13,28 +18,61 @@ BACKENDS = {
     "Qdrant - RAG": "Qdrant"
 }
 
-FAISS_PATH = "faiss_index_hnsw"
-QDRANT_PATH = "qdrant_db"
+# Resolve paths relative to this file so they work in HF Spaces
+_HERE = os.path.dirname(os.path.abspath(__file__))
+METADATA_PATH  = os.path.join(_HERE, "metadata.pkl")
+FAISS_PATH     = os.path.join(_HERE, "faiss_index_hnsw")
+QDRANT_PATH    = os.path.join(_HERE, "qdrant_db")
 QDRANT_COLLECTION = "articles"
 
-# ====================== LOAD METADATA FOR ARTICLE LIST ======================
-def load_articles_list():
+# ====================== LOAD METADATA ======================
+def _load_metadata():
+    """Load the DataFrame from metadata.pkl; return None on any failure."""
     try:
-        with open("metadata.pkl", "rb") as f:
+        with open(METADATA_PATH, "rb") as f:
             df = pickle.load(f)
-        articles = sorted(df['article_title'].unique().tolist())
-        return ["All articles"] + articles
-    except:
-        return ["All articles"]
+        logger.info("metadata.pkl loaded — %d rows, columns: %s", len(df), df.columns.tolist())
+        return df
+    except FileNotFoundError:
+        logger.error("metadata.pkl not found at %s", METADATA_PATH)
+    except Exception as exc:
+        logger.error("Failed to load metadata.pkl: %s", exc)
+    return None
+
+_METADATA_DF = _load_metadata()
+
+
+def load_category_list():
+    """Return ['All categories'] + sorted unique article_category values."""
+    if _METADATA_DF is not None and "article_category" in _METADATA_DF.columns:
+        cats = sorted(_METADATA_DF["article_category"].dropna().unique().tolist())
+        logger.info("Found %d categories", len(cats))
+        return ["All categories"] + cats
+    logger.warning("article_category column not found — showing only 'All categories'")
+    return ["All categories"]
+
+
+def load_articles_for_category(category: str):
+    """Return ['All articles in category'] + sorted titles for the given category."""
+    default = ["All articles in category"]
+    if _METADATA_DF is None or "article_title" not in _METADATA_DF.columns:
+        return default
+    if category in ("All categories", None, ""):
+        titles = sorted(_METADATA_DF["article_title"].dropna().unique().tolist())
+    else:
+        mask   = _METADATA_DF["article_category"] == category
+        titles = sorted(_METADATA_DF.loc[mask, "article_title"].dropna().unique().tolist())
+    return default + titles
 
-ARTICLE_LIST = load_articles_list()
+
+CATEGORY_LIST = load_category_list()
 
 # ====================== LOAD LLM ======================
 model_path = hf_hub_download(
     repo_id=repo_id,
     filename="articles-Q4_K_M.gguf",
     repo_type="model",
-    token=os.environ.get("HF_TOKEN")
+    token=os.environ.get("HF_TOKEN"),
 )
 
 llm = Llama(
@@ -47,14 +85,17 @@ llm = Llama(
 )
 
 # ====================== RAG CACHE ======================
-vectorstores = {}
+vectorstores: dict = {}
+
 
 def get_vectorstore(backend_name: str):
     if backend_name in vectorstores:
         return vectorstores[backend_name]
     try:
-        embeddings = HuggingFaceEmbeddings(model_name="BAAI/bge-small-en-v1.5", encode_kwargs={'normalize_embeddings': True})
-        
+        embeddings = HuggingFaceEmbeddings(
+            model_name="BAAI/bge-small-en-v1.5",
+            encode_kwargs={"normalize_embeddings": True},
+        )
         if backend_name == "FAISS":
             from langchain_community.vectorstores import FAISS
             vs = FAISS.load_local(FAISS_PATH, embeddings, allow_dangerous_deserialization=True)
@@ -64,31 +105,54 @@ def get_vectorstore(backend_name: str):
         else:
             from langchain_community.vectorstores import FAISS
             vs = FAISS.load_local(FAISS_PATH, embeddings, allow_dangerous_deserialization=True)
-            
         vectorstores[backend_name] = vs
+        logger.info("Vector store '%s' loaded successfully", backend_name)
         return vs
-    except:
+    except Exception as exc:
+        logger.error("Failed to load vector store '%s': %s", backend_name, exc)
         return None
 
 
 # ====================== SYSTEM PROMPT ======================
-SYSTEM_PROMPT = """You are the reference expert for the articles contained in the training of this model, all extracted from the website robertolofaro.com, and all focused on change.
-#Your Mission:
-When a user asks a question, your goal is to provide a structured response based ONLY on the articles provided in your training. Do not provide general advice from outside these sources.
-Do not provide article titles or article IDs, provide only the concepts that articles  express.
-# Response Format:
+# The explicit declaration that context is injected inline prevents the model from
+# reasoning "I have no access to the vector store" during its <think> block.
+SYSTEM_PROMPT = """You are the reference expert for the articles contained in the training \
+of this model, all extracted from the website robertolofaro.com, and all focused on change.
+
+IMPORTANT: Relevant article excerpts retrieved via semantic search will be injected \
+directly in the user message under the heading "Context:". You MUST use those excerpts \
+as the primary source for your answer. Do not speculate about whether you have access \
+to articles — the context IS provided inline when available.
+
+# Your Mission
+When a user asks a question, provide a structured response based ONLY on the article \
+content provided in the Context section. Do not draw on general knowledge outside those \
+sources. Do not provide article titles or article IDs — provide only the concepts the \
+articles express.
+
+# Response Format
 1. Executive Summary: A 2-3 sentence overview answering the core query.
-2. Guidelines & Hints: A markdown list of specific "answers/guidelines/hints" found in the source material.
-"""
+2. Guidelines & Hints: A markdown list of specific answers/guidelines/hints found in \
+the source material."""
 
 
 # ====================== GENERATION FUNCTION ======================
-def generate_response(message, history, rag_mode, article_filter, max_tokens, temperature, top_p, repeat_penalty):
+def generate_response(
+    message, history,
+    rag_mode, category_filter, article_filter,
+    max_tokens, temperature, top_p, repeat_penalty,
+    suppress_thinking,
+):
+    # Strip /nothink from the user-visible message if they typed it
+    clean_message = message.replace("/nothink", "").strip()
+
+    # Build prompt
     full_prompt = f"<|im_start|>system\n{SYSTEM_PROMPT}<|im_end|>\n"
 
     for msg in history[-4:]:
         full_prompt += f"<|im_start|>{msg['role']}\n{msg['content']}<|im_end|>\n"
 
+    # --- RAG retrieval ---
     backend = BACKENDS.get(rag_mode)
     context = ""
 
@@ -96,26 +160,44 @@ def generate_response(message, history, rag_mode, article_filter, max_tokens, te
         vs = get_vectorstore(backend)
         if vs:
             try:
-                filter_dict = {"article_title": article_filter} if article_filter != "All articles" else None
-                docs = vs.similarity_search(message, k=5, filter=filter_dict)
-                context = "\n\n".join([
-                    f"[Article: {doc.metadata.get('article_title', 'N/A')}] {doc.page_content[:700]}"
-                    for doc in docs
-                ])
-            except:
-                pass
+                # Build metadata filter: specific article takes priority over category
+                filter_dict = None
+                if article_filter and article_filter != "All articles in category":
+                    filter_dict = {"article_title": article_filter}
+                elif category_filter and category_filter != "All categories":
+                    filter_dict = {"article_category": category_filter}
+
+                docs = vs.similarity_search(clean_message, k=5, filter=filter_dict)
+                if docs:
+                    context = "\n\n".join(
+                        f"[Article: {doc.metadata.get('article_title', 'N/A')}] "
+                        f"{doc.page_content[:700]}"
+                        for doc in docs
+                    )
+                    logger.info("RAG retrieved %d chunks (filter=%s)", len(docs), filter_dict)
+                else:
+                    logger.warning("RAG returned 0 chunks for filter=%s", filter_dict)
+            except Exception as exc:
+                logger.error("RAG retrieval failed: %s", exc)
+
+    # Append /nothink to suppress Qwen3 thinking if requested
+    nothink_suffix = " /nothink" if suppress_thinking else ""
 
     if context:
-        full_prompt += f"<|im_start|>user\nContext:\n{context}\n\nQuestion: {message}<|im_end|>\n"
+        full_prompt += (
+            f"<|im_start|>user\nContext:\n{context}\n\n"
+            f"Question: {clean_message}{nothink_suffix}<|im_end|>\n"
+        )
     else:
-        full_prompt += f"<|im_start|>user\n{message}<|im_end|>\n"
+        full_prompt += f"<|im_start|>user\n{clean_message}{nothink_suffix}<|im_end|>\n"
 
     full_prompt += "<|im_start|>assistant\n"
 
-    max_tokens_val = int(max_tokens) if max_tokens is not None else 900
-    temp_val = float(temperature) if temperature is not None else 0.65
-    top_p_val = float(top_p) if top_p is not None else 0.9
-    rep_penalty_val = float(repeat_penalty) if repeat_penalty is not None else 1.1
+    # Sanitise generation params
+    max_tokens_val   = int(max_tokens)   if max_tokens   is not None else 900
+    temp_val         = float(temperature) if temperature  is not None else 0.65
+    top_p_val        = float(top_p)       if top_p        is not None else 0.9
+    rep_penalty_val  = float(repeat_penalty) if repeat_penalty is not None else 1.1
 
     partial_text = ""
     for chunk in llm(
@@ -127,10 +209,11 @@ def generate_response(message, history, rag_mode, article_filter, max_tokens, te
         stop=["<|im_end|>", "<|im_start|>"],
         stream=True,
     ):
-        token = chunk['choices'][0]['text']
+        token = chunk["choices"][0]["text"]
         partial_text += token
         yield partial_text
 
+
 # ====================== GRADIO INTERFACE ======================
 with gr.Blocks(title="Article Q&A model") as demo:
     gr.Markdown("# sourcing 350+ articles on change")
@@ -142,7 +225,8 @@ with gr.Blocks(title="Article Q&A model") as demo:
     gr.Markdown(
         "**NOTAM:** by querying this model you access the articles and metadata "
         "available on robertolofaro.com and GitHub.  "
-        "Answers reflect the article corpus only — do not treat them as advice- just expression of a position contained within the articles."
+        "Answers reflect the article corpus only — do not treat them as advice, "
+        "just expression of a position contained within the articles."
     )
     gr.Markdown(
         "If, after getting an answer, you want something tailored to your context, "
@@ -153,27 +237,56 @@ with gr.Blocks(title="Article Q&A model") as demo:
         rag_mode = gr.Radio(
             choices=list(BACKENDS.keys()),
             value="FAISS - RAG (HNSW)",
-            label="Mode"
+            label="Retrieval backend",
+        )
+        suppress_thinking = gr.Checkbox(
+            value=True,
+            label="Suppress model thinking (/nothink)",
+            info="Uncheck to see the model's reasoning chain",
+        )
+
+    with gr.Row():
+        category_filter = gr.Dropdown(
+            choices=CATEGORY_LIST,
+            value="All categories",
+            label="Filter by category",
+            info=f"{len(CATEGORY_LIST) - 1} categories available",
         )
         article_filter = gr.Dropdown(
-            choices=ARTICLE_LIST,
-            value="All articles",
-            label="Focus on article"
+            choices=["All articles in category"],
+            value="All articles in category",
+            label="Narrow to specific article (optional)",
+            info="Select a category first to populate this list",
         )
 
+    # Dynamically populate the article dropdown when category changes
+    def update_article_dropdown(category):
+        articles = load_articles_for_category(category)
+        return gr.Dropdown(choices=articles, value=articles[0])
+
+    category_filter.change(
+        fn=update_article_dropdown,
+        inputs=category_filter,
+        outputs=article_filter,
+    )
+
     with gr.Accordion("Advanced Generation Parameters", open=False):
-        max_tokens = gr.Slider(256, 2048, value=900, step=64, label="Max Tokens")
-        temperature = gr.Slider(0.0, 1.0, value=0.65, step=0.05, label="Temperature")
-        top_p = gr.Slider(0.0, 1.0, value=0.9, step=0.05, label="Top-p")
-        repeat_penalty = gr.Slider(1.0, 2.0, value=1.1, step=0.05, label="Repeat Penalty")
+        max_tokens    = gr.Slider(256, 2048, value=900,  step=64,   label="Max Tokens")
+        temperature   = gr.Slider(0.0,  1.0, value=0.65, step=0.05, label="Temperature")
+        top_p         = gr.Slider(0.0,  1.0, value=0.9,  step=0.05, label="Top-p")
+        repeat_penalty = gr.Slider(1.0, 2.0, value=1.1,  step=0.05, label="Repeat Penalty")
 
     gr.ChatInterface(
         fn=generate_response,
-        additional_inputs=[rag_mode, article_filter, max_tokens, temperature, top_p, repeat_penalty],
-        cache_examples=False, # <--- Stops Gradio from executing them at startup
+        additional_inputs=[
+            rag_mode, category_filter, article_filter,
+            max_tokens, temperature, top_p, repeat_penalty,
+            suppress_thinking,
+        ],
+        cache_examples=False,
         examples=[
-            ["What is the potential for Italy? /nothink"],
-            ["What is the potential for Turin? /nothink"]
+            ["What is the potential for Italy?"],
+            ["What is the potential for Turin?"],
         ],
     )