Deploy 13e6fc0

app/pipeline/nodes/generate.py

@@ -1,5 +1,4 @@
 from typing import Callable
-import re
 
 from app.models.pipeline import PipelineState
 from app.models.chat import SourceRef
@@ -40,12 +39,14 @@ def make_generate_node(llm_client: LLMClient) -> Callable[[PipelineState], dict]
         context_block = "\n\n".join(context_parts)
         
         system_prompt = (
-            "You are Darshan's personal assistant. Answer using ONLY the provided "
-            "context. After each factual claim, cite inline as [Title](url). "
-            "Be concise and direct. Never invent information not in the context."
+            "You are Darshan Chheda's personal AI assistant embedded on his portfolio. "
+            "Answer questions using ONLY the numbered context passages below. "
+            "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."
         )
-        
-        prompt = f"Context:\n{context_block}\n\nQuery: {query}"
+
+        prompt = f"Context:\n{context_block}\n\nQuestion: {query}"
         
         # Complete via the requested streams
         stream = llm_client.complete_with_complexity(prompt=prompt, system=system_prompt, stream=True, complexity=complexity)
@@ -53,28 +54,13 @@ def make_generate_node(llm_client: LLMClient) -> Callable[[PipelineState], dict]
         full_answer = ""
         async for chunk in stream:
              full_answer += chunk
-             
-        # Extract source refs mentioned in the response
-        # Matches markdown links like [Title](url)
-        mentioned_refs = []
-        
-        # Regex to find all [Text](URL)
-        matches = re.findall(r"\[([^\]]+)\]\(([^)]+)\)", full_answer)
-        # Find which of our source refs match the URL
-        url_to_ref = {ref.url: ref for ref in source_refs}
-        
-        for title, url in matches:
-             if url in url_to_ref:
-                  if url_to_ref[url] not in mentioned_refs:
-                      mentioned_refs.append(url_to_ref[url])
-                      
-        # Fallback: if no specific inline citations were used but we have sources,
-        # we can attach all provided sources, or strictly just the mentioned ones.
-        # "extracts source refs mentioned in response" -> we return `mentioned_refs`
-        
+
+        # Always return all source refs used as context — the LLM is instructed
+        # to cite inline as [N], so every chunk in context is a potential citation.
+        # Filtering by regex is fragile; the frontend renders all sources as footnotes.
         return {
              "answer": full_answer,
-             "sources": mentioned_refs
+             "sources": source_refs
         }
 
     return generate_node

app/pipeline/nodes/retrieve.py

@@ -30,10 +30,11 @@ def make_retrieve_node(vector_store: VectorStore, embedder: Embedder, reranker:
         # Reranker is async — must be awaited.
         reranked = await reranker.rerank(state["query"], unique_chunks, top_k=5)
 
-        # Low-confidence fallback: skip generation and return early.
-        if not reranked or reranker.min_score < 0.45:
+        # No chunks at all: collection is empty or query is too niche.
+        # Return empty so generate node returns its fallback cleanly.
+        if not reranked:
             return {
-                "answer": "I don't have enough information about this in my knowledge base. Try asking about my specific projects or blog posts.",
+                "answer": "I don't have enough information about this in my knowledge base. Try asking about Darshan's specific projects or blog posts.",
                 "retrieved_chunks": [],
                 "reranked_chunks": [],
             }