Deploy cf7f476

app/pipeline/nodes/generate.py

@@ -40,14 +40,14 @@ def make_generate_node(llm_client: LLMClient) -> Callable[[PipelineState], dict]
         context_block = "\n\n".join(context_parts)
         
         system_prompt = (
-            "You are an AI assistant embedded on Darshan Chheda's personal portfolio website. "
-            "The numbered context passages below describe Darshan's projects, blog posts, skills, and experiences. "
-            "Darshan Chheda is the subject — always refer to him by name or as 'he'/'his'. "
-            "When a passage uses 'I', 'my', or 'me', that voice belongs to Darshan. "
-            "Answer questions using ONLY the context passages provided. "
-            "Cite sources inline using bracketed numbers like [1], [2] immediately after each claim. "
-            "Be concise, confident, and factual. Never invent details not present in the context. "
-            "If the context doesn't contain enough information to answer fully, say so honestly."
+            "You are a knowledgeable assistant on Darshan Chheda's personal portfolio website. "
+            "Use the numbered context passages below — which cover Darshan's projects, blog posts, skills, and background — to answer the question. "
+            "Darshan Chheda is always the subject. When a passage says 'I', 'my', or 'me', that voice is Darshan's. "
+            "Write in clear, professional English as if you know Darshan well. "
+            "You may draw reasonable, confident inferences from the evidence (e.g. if he built a full-stack app he knows both front-end and back-end; if he wrote Assembly he understands low-level programming). "
+            "Do NOT say you cannot find information when the context clearly contains the answer — read it carefully. "
+            "Cite every factual claim inline with bracketed numbers like [1], [2] immediately after the claim. "
+            "Be concise and direct. Avoid hedging phrases like 'based on the context' or 'the context suggests'."
         )
 
         prompt = f"Context:\n{context_block}\n\nQuestion: {query}"

app/security/sanitizer.py

@@ -59,7 +59,10 @@ def redact_pii(text: str) -> str:
          # Failsafe if Presidio isn't installed/working
          return text
          
-    entities = ["EMAIL_ADDRESS", "PHONE_NUMBER", "UK_NHS", "IBAN_CODE", "PERSON"]
+    # PERSON is intentionally excluded: visitors are expected to name Darshan Chheda
+    # in their queries. Redacting that breaks retrieval and confuses the LLM.
+    # We only protect against visitor PII that could leak into logs (e-mail, phone, etc.).
+    entities = ["EMAIL_ADDRESS", "PHONE_NUMBER", "UK_NHS", "IBAN_CODE"]
     
     try:
         results = analyzer.analyze(text=text, entities=entities, language='en')