Deploy 0c3a28c

app/pipeline/nodes/generate.py

@@ -1,3 +1,4 @@
+import re
 from typing import Callable
 
 from app.models.pipeline import PipelineState
@@ -39,8 +40,11 @@ def make_generate_node(llm_client: LLMClient) -> Callable[[PipelineState], dict]
         context_block = "\n\n".join(context_parts)
         
         system_prompt = (
-            "You are Darshan Chheda's personal AI assistant embedded on his portfolio. "
-            "Answer questions using ONLY the numbered context passages below. "
+            "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."
@@ -55,12 +59,16 @@ def make_generate_node(llm_client: LLMClient) -> Callable[[PipelineState], dict]
         async for chunk in stream:
              full_answer += chunk
 
-        # 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.
+        # Only surface source refs that the LLM actually cited with [N] markers.
+        # Returning all context chunks floods the frontend with irrelevant footnotes.
+        cited_indices = {int(m) for m in re.findall(r"\[(\d+)\]", full_answer)}
+        cited_sources = [
+            sr for i, sr in enumerate(source_refs, start=1) if i in cited_indices
+        ]
+
         return {
              "answer": full_answer,
-             "sources": source_refs
+             "sources": cited_sources if cited_sources else source_refs[:2]
         }
 
     return generate_node

app/services/vector_store.py

@@ -2,7 +2,7 @@ import uuid
 from typing import Optional
 
 from qdrant_client import QdrantClient
-from qdrant_client.models import PointStruct, VectorParams, Distance, Filter, FieldCondition, MatchValue
+from qdrant_client.models import PointStruct, VectorParams, Distance, Filter, FieldCondition, MatchValue, PayloadSchemaType
 
 from app.models.pipeline import Chunk, ChunkMetadata
 from app.core.exceptions import RetrievalError
@@ -14,16 +14,25 @@ class VectorStore:
         self.collection = collection
 
     def ensure_collection(self) -> None:
-        """Creates collection with vectors size=384, distance=Cosine if it does not exist."""
+        """Creates collection with vectors size=384, distance=Cosine if it does not exist.
+        Also ensures payload index on metadata.doc_id exists for efficient dedup deletes."""
         collections = self.client.get_collections().collections
         exists = any(c.name == self.collection for c in collections)
-        
+
         if not exists:
             self.client.create_collection(
                 collection_name=self.collection,
                 vectors_config=VectorParams(size=384, distance=Distance.COSINE),
             )
 
+        # Keyword index allows filter-by-doc_id in delete_by_doc_id.
+        # create_payload_index is idempotent — safe to call on every startup.
+        self.client.create_payload_index(
+            collection_name=self.collection,
+            field_name="metadata.doc_id",
+            field_schema=PayloadSchemaType.KEYWORD,
+        )
+
     def upsert_chunks(self, chunks: list[Chunk], embeddings: list[list[float]]) -> None:
         """Builds PointStruct list and calls client.upsert. Batch size 100."""
         if len(chunks) != len(embeddings):