Deploy 33c771f

app/api/chat.py

@@ -246,6 +246,7 @@ async def chat_endpoint(
         "interaction_id": None,
         "retrieval_attempts": 0,
         "rewritten_query": None,
+        "rewrite_changed": False,
         "follow_ups": [],
         "path": None,
         "query_topic": None,

app/models/pipeline.py

@@ -90,6 +90,9 @@ class PipelineState(TypedDict):
     # CRAG: counts retrieve node invocations; 2 = one retry was attempted.
     retrieval_attempts: int
     rewritten_query: Optional[str]
+    # Whether the last CRAG rewrite produced a different query; a retry with the same one
+    # can only find the same chunks.
+    rewrite_changed: bool
     # Top cross-encoder score from the last retrieve call. Used by CRAG routing.
     top_rerank_score: Optional[float]
     follow_ups: list[str]

app/pipeline/graph.py

@@ -116,6 +116,18 @@ def route_retrieve_result(state: PipelineState) -> str:
     return "generate"
 
 
+def route_rewrite(state: PipelineState) -> str:
+    """
+    Retry retrieval with a different query, or when nothing usable was found (a retry
+    then widens to the broader fallback). Otherwise, when the rewrite could not change the
+    query (Gemini not configured, failed, or returned it as it was), a second retrieve finds
+    the same chunks and only adds its latency, so the chunks already found are answered from.
+    """
+    if state.get("rewrite_changed", False) or not state.get("reranked_chunks"):
+        return "retry"
+    return "generate"
+
+
 def build_pipeline(services: dict) -> CompiledStateGraph:
     graph = StateGraph(PipelineState)
 
@@ -155,7 +167,8 @@ def build_pipeline(services: dict) -> CompiledStateGraph:
 
     # After rewrite: go straight back to retrieve for the next attempt.
     # The cycle terminates because route_retrieve_result checks retrieval_attempts.
-    graph.add_edge("rewrite_query", "retrieve")
+    graph.add_conditional_edges("rewrite_query", route_rewrite,
+                                {"retry": "retrieve", "generate": "generate"})
 
     graph.add_edge("generate", "log_eval")
     graph.add_edge("log_eval", END)

app/pipeline/nodes/rewrite_query.py

@@ -67,6 +67,7 @@ def make_rewrite_query_node(gemini_client: GeminiClient) -> Any:
             logger.debug("Gemini not configured; skipping query rewrite.")
             return {
                 "rewritten_query": query,
+                "rewrite_changed": False,
                 "retrieval_attempts": state.get("retrieval_attempts", 1) + 1,
                 "query_embedding": None,  # Force re-embed so retrieve doesn't use stale embedding
             }
@@ -94,6 +95,7 @@ def make_rewrite_query_node(gemini_client: GeminiClient) -> Any:
         return {
             "query": rewritten,
             "rewritten_query": rewritten,
+            "rewrite_changed": rewritten != query,
             "retrieval_attempts": state.get("retrieval_attempts", 1) + 1,
             "query_embedding": None,
         }

tests/test_graph_routing.py

@@ -28,3 +28,13 @@ def test_attempt_three_unrelated_low_confidence_generates() -> None:
         "query": "what is the weather in london",
     }
     assert route_retrieve_result(state) == "generate"
+
+
+def test_an_unchanged_rewrite_answers_from_the_chunks_already_found() -> None:
+    from app.pipeline.graph import route_rewrite
+
+    found = [{"text": "x", "metadata": {}}]
+    assert route_rewrite({"rewrite_changed": False, "reranked_chunks": found}) == "generate"
+    assert route_rewrite({"rewrite_changed": True, "reranked_chunks": found}) == "retry"
+    # Nothing usable yet: the retry is what widens to the broader fallback.
+    assert route_rewrite({"rewrite_changed": False, "reranked_chunks": []}) == "retry"

tests/test_site_profile.py

@@ -5,7 +5,7 @@ ROOT = Path(__file__).resolve().parents[2]
 if str(ROOT) not in sys.path:
     sys.path.insert(0, str(ROOT))
 
-from ingestion.site_profile import profile_documents
+from ingestion.site_profile import profile_chunks
 
 LLMS_TXT = """# Darshan Chheda
 
@@ -25,14 +25,23 @@ Open to software engineering roles in the UK
 """
 
 
-def test_the_page_is_one_profile_and_each_off_site_profile_its_own_document() -> None:
-    documents = profile_documents(LLMS_TXT)
+def test_each_section_is_a_chunk_that_says_whose_section_it_is() -> None:
+    [(page, label), *_] = profile_chunks(LLMS_TXT)
 
-    assert [(d["title"], d["url"], d["slug"], d["source_type"]) for d in documents] == [
-        ("darshanchheda.com", "https://www.darshanchheda.com/", "site-profile", "profile"),
-        ("GitHub", "https://github.com/1337XCode", "profile-github", "contact"),
-        ("LinkedIn", "https://linkedin.com/in/darshchheda", "profile-linkedin", "contact"),
+    assert label == "site-profile"
+    assert [chunk["metadata"]["section"] for chunk in page] == ["Profile", "Availability", "Links", "Pages"]
+    assert page[0]["text"] == "Darshan Chheda — Profile:\n> London-based software engineer."
+    assert page[1]["text"] == "Darshan Chheda — Availability:\nOpen to software engineering roles in the UK"
+    assert {chunk["metadata"]["source_type"] for chunk in page} == {"profile"}
+
+
+def test_each_off_site_profile_is_a_contact_chunk_of_its_own() -> None:
+    [_, *profiles] = profile_chunks(LLMS_TXT)
+
+    assert [(chunks[0]["metadata"]["source_title"], chunks[0]["metadata"]["source_url"], chunks[0]["metadata"]["source_type"]) for chunks, _ in profiles] == [
+        ("GitHub", "https://github.com/1337XCode", "contact"),
+        ("LinkedIn", "https://linkedin.com/in/darshchheda", "contact"),
     ]
-    assert "Open to software engineering roles in the UK" in documents[0]["clean_content"]
-    assert documents[2]["clean_content"].endswith("https://linkedin.com/in/darshchheda")
-    assert len(documents[2]["clean_content"].split()) >= 8
+    linkedin = profiles[1][0][0]["text"]
+    assert linkedin.endswith("https://linkedin.com/in/darshchheda")
+    assert len(linkedin.split()) >= 8