chore: sync app/ and src/ from GitHub

app/app.py

@@ -126,9 +126,10 @@ hybrid_retriever = HybridRetriever(
         semantic_weight=0.5,
     )
 
+
 def llm_retriever(query: str, top_k: int = 5):
-    answer, docs = run_rag(hybrid_retriever, query=query)
-    return answer, docs
+    answer, docs, web_sources = run_rag(hybrid_retriever, query=query)
+    return answer, docs, web_sources
 
 
 # ─── Helpers ──────────────────────────────────────────────────────────────────
@@ -272,12 +273,14 @@ if query.strip() and query != st.session_state.get("last_query"):
 
     with st.spinner("Asking AI..."):
         try:
-            answer, docs = llm_retriever(query, top_k=TOP_K)
+            answer, docs, web_sources = llm_retriever(query, top_k=TOP_K)
             st.session_state.llm_result = answer
             st.session_state.llm_docs = docs
+            st.session_state.web_sources = web_sources
         except Exception as e:
             st.session_state.llm_result = f"**Error:** {e}"
             st.session_state.llm_docs = []
+            st.session_state.web_sources = []
 
 elif not query.strip():
     # Clear results when input is emptied
@@ -338,6 +341,13 @@ with tab_llm:
         else: 
             st.markdown("<p style='color:#aaa;'>No documents retrieved.</p>", unsafe_allow_html=True)
 
+        # ── Web sources ───────────────────────────────────────────────────────
+        sources = st.session_state.get("web_sources", [])
+        if sources:
+            st.markdown("#### 🌐 Web Sources")
+            for s in sources:
+                st.markdown(f"- [{s['title']}]({s['url']})")
+
 # ─── Sidebar: feedback log ────────────────────────────────────────────────────
 with st.sidebar:
     st.header("📋 Feedback Log")

src/rag_pipeline.py

@@ -20,7 +20,9 @@ from langchain_core.documents import Document
 from langchain_core.output_parsers import StrOutputParser
 from langchain_core.prompts import ChatPromptTemplate
 from langchain_core.runnables import RunnableLambda, RunnablePassthrough
+import os
 from langchain_huggingface import ChatHuggingFace, HuggingFaceEndpoint
+
 # ---------------------------------------------------------------------------
 # Logging
 # ---------------------------------------------------------------------------
@@ -36,6 +38,12 @@ DEFAULT_TOP_K = 5
 DEFAULT_SYSTEM_PROMPT = (
     "You are a helpful Amazon grocery shopping assistant.\n\n"
     "You will receive a grocery query and a list of related Amazon products (including reviews and metadata).\n\n"
+
+    "If the context contains a section starting with 'Web search results', "
+    "incorporate that pricing or availability information naturally into your answer — "
+    "do not copy it verbatim or list raw numbers. Sources will be displayed separately, "
+    "so you do not need to include URLs in your response.\n\n"
+
     "Your response must follow this exact structure:\n\n"
     "---\n\n"
     "## 🛒 Recommended Products\n"
@@ -51,65 +59,62 @@ DEFAULT_SYSTEM_PROMPT = (
     "- Keep descriptions factual and grounded in the provided reviews and metadata.\n"
     "- Recipe ideas should be suggestions or ideas only, not step-by-step instructions.\n"
     "- Format the entire response in Markdown.\n"
+    "- If any information comes from a web search, cite the source inline as [source](url).\n"
     "- IMPORTANT: Whenever citing the product title: add the parent_asin in the following format [title](#parent_asin)"
 )
 
 # ---------------------------------------------------------------------------
 # Helper functions
 # ---------------------------------------------------------------------------
-
-import logging
 from langchain_core.runnables import RunnableLambda
 
-logger = logging.getLogger(__name__)
+# Keyword triggers that suggest the query needs external/current information
+_WEB_SEARCH_TRIGGERS = {
+    "price", "cost", "available", "availability", "recall", "news",
+    "latest", "current", "today", "recently", "substitute", "substitution",
+    "allergen", "gluten", "vegan", "organic", "nutrition", "calories",
+}
 
-def _make_verbose_tap(label: str, verbose: bool):
+def _maybe_web_search(query: str) -> tuple[str, list[dict]]:
     """
-    Returns a passthrough RunnableLambda that logs *value* when verbose=True.
-    Works for any chain step — docs, prompt messages, or raw strings.
+    Returns (context_string, sources_list) where sources_list is
+    [{"title": ..., "url": ...}, ...] for clean rendering.
     """
+    tokens = set(query.lower().split())
+    if tokens & _WEB_SEARCH_TRIGGERS:
+        try:
+            from tavily import TavilyClient
+            client = TavilyClient(api_key=os.getenv("TAVILY_API_KEY"))
+            response = client.search(query, max_results=3)
+            results = response.get("results", [])
+            snippets = "\n\n".join(r["content"] for r in results)
+            sources = [{"title": r.get("title", r["url"]), "url": r["url"]} for r in results]
+            context = f"\n\nWeb search results (use this to answer pricing/availability questions):\n{snippets}"
+            return context, sources
+        except Exception as e:
+            logger.warning("Web search failed: %s", e)
+    return "", []
+
+
+def _make_verbose_tap(label: str, verbose: bool):
     def _tap(value):
         if verbose:
-            if hasattr(value, "messages"):          # ChatPromptValue
+            if hasattr(value, "messages"):
                 rendered = "\n".join(
                     f"[{m.type.upper()}]: {m.content}"
                     for m in value.messages
                 )
-            elif isinstance(value, list):            # list of Documents
+            elif isinstance(value, list):
                 rendered = "\n".join(str(d) for d in value)
             else:
                 rendered = str(value)
-
             print(f"\n{'='*60}\n{label}\n{'='*60}\n{rendered}\n")
             logger.debug("%s\n%s", label, rendered)
         return value
     return RunnableLambda(_tap)
 
-def build_context(docs: list[Document]) -> str:
-    """
-    Concatenate a list of retrieved LangChain Documents into a single
-    context string that the LLM can reason over.
-
-    Each entry includes the product's ``parent_asin`` (falling back to its
-    position index), its page content, and its full metadata dict.
-
-    Parameters
-    ----------
-    docs:
-        List of ``langchain_core.documents.Document`` objects returned by
-        the retriever.
 
-    Returns
-    -------
-    str
-        A newline-separated block of product descriptions ready for prompt
-        injection. Returns an empty string when *docs* is empty.
-
-    Raises
-    ------
-    TypeError
-        If *docs* is not a list, or any element is not a ``Document``.
-    """
+def build_context(docs: list[Document]) -> str:
     if not isinstance(docs, list):
         raise TypeError(
             f"'docs' must be a list of Document objects, got {type(docs).__name__}."
@@ -119,11 +124,9 @@ def build_context(docs: list[Document]) -> str:
             raise TypeError(
                 f"Element at index {i} is not a Document; got {type(doc).__name__}."
             )
-
     if not docs:
         logger.warning("build_context received an empty document list.")
         return ""
-
     return "\n\n".join(
         f"ASIN {doc.metadata.get('parent_asin', n)} Description: {doc.page_content}\n"
         f"Metadata: {doc.metadata}"
@@ -136,26 +139,6 @@ def _build_llm(
     max_new_tokens: int,
     provider: str,
 ) -> ChatHuggingFace:
-    """
-    Instantiate and return a ``ChatHuggingFace`` model backed by a
-    HuggingFace Inference Endpoint.
-
-    Parameters
-    ----------
-    repo_id:
-        HuggingFace Hub model identifier (e.g.
-        ``"meta-llama/Meta-Llama-3-8B-Instruct"``).
-    max_new_tokens:
-        Maximum number of tokens the model may generate per call.
-    provider:
-        Inference provider passed to ``HuggingFaceEndpoint``
-        (``"auto"``, ``"novita"``, etc.).
-
-    Returns
-    -------
-    ChatHuggingFace
-        A chat-compatible wrapper around the endpoint.
-    """
     endpoint = HuggingFaceEndpoint(
         repo_id=repo_id,
         task="text-generation",
@@ -166,19 +149,6 @@ def _build_llm(
 
 
 def _build_prompt_template(system_prompt: str) -> ChatPromptTemplate:
-    """
-    Create a ``ChatPromptTemplate`` with a system message and a human
-    turn that injects ``{context}`` and ``{question}`` placeholders.
-
-    Parameters
-    ----------
-    system_prompt:
-        The system-level instruction string.
-
-    Returns
-    -------
-    ChatPromptTemplate
-    """
     return ChatPromptTemplate.from_messages([
         ("system", system_prompt),
         (
@@ -201,87 +171,34 @@ def run_rag(
     max_new_tokens: int = DEFAULT_MAX_NEW_TOKENS,
     provider: str = "auto",
     verbose: bool = False,
-) -> str:
-    """
-    Execute a full RAG pipeline and return the model's answer.
-
-    The pipeline follows the steps below:
-
-    1. **Retrieve** - *retriever* fetches the *k* most relevant documents
-       for *query*.
-    2. **Format context** - :func:`build_context` serialises the documents
-       into a single string.
-    3. **Prompt** - the context and query are injected into the chat prompt
-       template.
-    4. **Generate** - the LLM produces an answer grounded in the context.
-    5. **Parse** - the raw chat message is unwrapped to a plain string.
-
-    Parameters
-    ----------
-    retriever:
-        A LangChain-compatible retriever (must expose ``.invoke()`` and be
-        pipeable with ``|``).  Typically created via
-        ``vectorstore.as_retriever(...)``.
-    query:
-        Natural-language question to answer (non-empty string).
-    system_prompt:
-        System-level instruction for the assistant.  Defaults to
-        :data:`DEFAULT_SYSTEM_PROMPT`.
-    repo_id:
-        HuggingFace Hub model identifier.  Defaults to
-        ``"meta-llama/Meta-Llama-3-8B-Instruct"``.
-    max_new_tokens:
-        Upper bound on generated tokens.  Must be a positive integer.
-        Defaults to ``100``.
-    provider:
-        HuggingFace inference provider (e.g. ``"auto"``, ``"novita"``).
-        Defaults to ``"auto"``.
-
-    Returns
-    -------
-    str
-        The model's answer as a plain string.
-
-    Raises
-    ------
-    TypeError
-        If *retriever* is ``None``, *query* is not a string, or
-        *system_prompt* is not a string.
-    ValueError
-        If *query* is blank, *max_new_tokens* is not a positive integer,
-        or *repo_id* / *provider* are blank strings.
-
-    Examples
-    --------
-    >>> answer = run_rag(retriever, "Best waterproof mascara under $20")
-    >>> print(answer)
-    """
+) -> tuple[str, list[Document]]:
     # ------------------------------------------------------------------
     # Build chain components
     # ------------------------------------------------------------------
-
     logger.info("Initialising LLM endpoint: %s", repo_id)
     llm = _build_llm(repo_id, max_new_tokens, provider)
     prompt_template = _build_prompt_template(system_prompt)
 
-    retrieved_docs: list[Document] = []   # ← capture target
+    web_context, web_sources = _maybe_web_search(query)
+
+    retrieved_docs: list[Document] = []
 
     def _retrieve_and_capture(query: str) -> list[Document]:
-        """Invoke the retriever and snapshot the results for the caller."""
         docs = retriever.invoke(query)
-        retrieved_docs.extend(docs)       # ← populate closure variable
-        return docs                       # ← pass through to build_context
+        retrieved_docs.extend(docs)
+        return docs
 
     rag_chain = (
         {
             "context": RunnableLambda(_retrieve_and_capture)
                        | RunnableLambda(build_context)
+                       | RunnableLambda(lambda ctx: ctx + web_context)
                        | _make_verbose_tap("RETRIEVED CONTEXT", verbose),
             "question": RunnablePassthrough(),
         }
         | _make_verbose_tap("PROMPT INPUTS (context + question)", verbose)
         | prompt_template
-        | _make_verbose_tap("RENDERED PROMPT SENT TO LLM", verbose)   # ← shows exact prompt
+        | _make_verbose_tap("RENDERED PROMPT SENT TO LLM", verbose)
         | llm
         | StrOutputParser()
     )
@@ -293,4 +210,4 @@ def run_rag(
     answer: str = rag_chain.invoke(query)
     logger.debug("RAG answer: %s", answer)
 
-    return answer, retrieved_docs
+    return answer, retrieved_docs, web_sources

src/tools.py

@@ -0,0 +1,17 @@
+import os
+from langchain.tools import tool
+
+@tool
+def web_search(query: str, max_results: int = 3) -> str:
+    """
+    Search the web for current information about a grocery or gourmet food product.
+    Use this when the user asks about recent news, current pricing, availability,
+    updated nutritional info, or anything unlikely to be in the product review corpus.
+    Input should be a specific product name or question.
+    """
+
+    from tavily import TavilyClient
+    client = TavilyClient(api_key=os.getenv("TAVILY_API_KEY"))
+    results = client.search(query, max_results=max_results)
+    snippets = [r["content"] for r in results.get("results", [])]
+    return "\n\n".join(snippets) if snippets else "No results found."