Add answers to search (#2020)

backend/ee/danswer/server/query_and_chat/models.py

@@ -1,6 +1,7 @@
 from pydantic import BaseModel
 
 from danswer.configs.constants import DocumentSource
+from danswer.search.enums import LLMEvaluationType
 from danswer.search.enums import SearchType
 from danswer.search.models import ChunkContext
 from danswer.search.models import RetrievalDetails
@@ -21,9 +22,9 @@ class DocumentSearchRequest(ChunkContext):
     search_type: SearchType
     retrieval_options: RetrievalDetails
     recency_bias_multiplier: float = 1.0
+    evaluation_type: LLMEvaluationType
     # This is to forcibly skip (or run) the step, if None it uses the system defaults
     skip_rerank: bool | None = None
-    skip_llm_chunk_filter: bool | None = None
 
 
 class BasicCreateChatMessageRequest(ChunkContext):

backend/ee/danswer/server/query_and_chat/query_backend.py

@@ -1,3 +1,5 @@
+from typing import cast
+
 from fastapi import APIRouter
 from fastapi import Depends
 from fastapi import HTTPException
@@ -9,7 +11,9 @@ from danswer.configs.danswerbot_configs import DANSWER_BOT_TARGET_CHUNK_PERCENTA
 from danswer.danswerbot.slack.handlers.handle_standard_answers import (
     oneoff_standard_answers,
 )
+from danswer.db.chat import translate_db_search_doc_to_server_search_doc
 from danswer.db.engine import get_session
+from danswer.db.models import SearchDoc
 from danswer.db.models import User
 from danswer.db.persona import get_persona_by_id
 from danswer.llm.answering.prompts.citations_prompt import (
@@ -27,6 +31,7 @@ from danswer.search.models import SearchRequest
 from danswer.search.pipeline import SearchPipeline
 from danswer.search.utils import dedupe_documents
 from danswer.search.utils import drop_llm_indices
+from danswer.search.utils import relevant_documents_to_indices
 from danswer.utils.logger import setup_logger
 from ee.danswer.server.query_and_chat.models import DocumentSearchRequest
 from ee.danswer.server.query_and_chat.models import StandardAnswerRequest
@@ -63,7 +68,7 @@ def handle_search_request(
             offset=search_request.retrieval_options.offset,
             limit=search_request.retrieval_options.limit,
             skip_rerank=search_request.skip_rerank,
-            skip_llm_chunk_filter=search_request.skip_llm_chunk_filter,
+            evaluation_type=search_request.evaluation_type,
             chunks_above=search_request.chunks_above,
             chunks_below=search_request.chunks_below,
             full_doc=search_request.full_doc,
@@ -75,8 +80,7 @@ def handle_search_request(
         bypass_acl=False,
     )
     top_sections = search_pipeline.reranked_sections
-    # If using surrounding context or full doc, this will be empty
-    relevant_section_indices = search_pipeline.relevant_section_indices
+    relevance_sections = search_pipeline.section_relevance
     top_docs = [
         SavedSearchDocWithContent(
             document_id=section.center_chunk.document_id,
@@ -105,19 +109,26 @@ def handle_search_request(
     # Deduping happens at the last step to avoid harming quality by dropping content early on
     deduped_docs = top_docs
     dropped_inds = None
+
     if search_request.retrieval_options.dedupe_docs:
         deduped_docs, dropped_inds = dedupe_documents(top_docs)
 
+    llm_indices = relevant_documents_to_indices(
+        relevance_sections=relevance_sections,
+        search_docs=[
+            translate_db_search_doc_to_server_search_doc(cast(SearchDoc, doc))
+            for doc in deduped_docs
+        ],
+    )
+
     if dropped_inds:
-        relevant_section_indices = drop_llm_indices(
-            llm_indices=relevant_section_indices,
+        llm_indices = drop_llm_indices(
+            llm_indices=llm_indices,
             search_docs=deduped_docs,
             dropped_indices=dropped_inds,
         )
 
-    return DocumentSearchResponse(
-        top_documents=deduped_docs, llm_indices=relevant_section_indices
-    )
+    return DocumentSearchResponse(top_documents=deduped_docs, llm_indices=llm_indices)
 
 
 @basic_router.post("/answer-with-quote")