ensure consistency of answers + update llm relevance prompting (#2045)

backend/danswer/prompts/llm_chunk_filter.py

@@ -11,7 +11,13 @@ it must contain information that is USEFUL for answering the query.
 If the section contains ANY useful information, that is good enough, \
 it does not need to fully answer the every part of the user query.
 
+
+Title: {{title}}
+
+{{optional_metadata}}
+
 Reference Section:
+
 ```
 {{chunk_text}}
 ```

backend/danswer/search/postprocessing/postprocessing.py

@@ -212,11 +212,17 @@ def filter_sections(
         section.center_chunk.content if use_chunk else section.combined_content
         for section in sections_to_filter
     ]
+    metadata_list = [section.center_chunk.metadata for section in sections_to_filter]
+    titles = [
+        section.center_chunk.semantic_identifier for section in sections_to_filter
+    ]
 
     llm_chunk_selection = llm_batch_eval_sections(
         query=query.query,
         section_contents=contents,
         llm=llm,
+        titles=titles,
+        metadata_list=metadata_list,
     )
 
     return [

backend/danswer/secondary_llm_flows/chunk_usefulness.py

@@ -12,17 +12,33 @@ from danswer.utils.threadpool_concurrency import run_functions_tuples_in_paralle
 logger = setup_logger()
 
 
-def llm_eval_section(query: str, section_content: str, llm: LLM) -> bool:
+def llm_eval_section(
+    query: str,
+    section_content: str,
+    llm: LLM,
+    title: str,
+    metadata: dict[str, str | list[str]],
+) -> bool:
+    def _get_metadata_str(metadata: dict[str, str | list[str]]) -> str:
+        metadata_str = "\n\nMetadata:\n"
+        for key, value in metadata.items():
+            value_str = ", ".join(value) if isinstance(value, list) else value
+            metadata_str += f"{key} - {value_str}\n"
+        return metadata_str + "\nContent:"
+
     def _get_usefulness_messages() -> list[dict[str, str]]:
+        metadata_str = _get_metadata_str(metadata) if metadata else ""
         messages = [
             {
                 "role": "user",
                 "content": SECTION_FILTER_PROMPT.format(
-                    chunk_text=section_content, user_query=query
+                    title=title,
+                    chunk_text=section_content,
+                    user_query=query,
+                    optional_metadata=metadata_str,
                 ),
             },
         ]
-
         return messages
 
     def _extract_usefulness(model_output: str) -> bool:
@@ -34,9 +50,6 @@ def llm_eval_section(query: str, section_content: str, llm: LLM) -> bool:
 
     messages = _get_usefulness_messages()
     filled_llm_prompt = dict_based_prompt_to_langchain_prompt(messages)
-    # When running in a batch, it takes as long as the longest thread
-    # And when running a large batch, one may fail and take the whole timeout
-    # instead cap it to 5 seconds
     model_output = message_to_string(llm.invoke(filled_llm_prompt))
     logger.debug(model_output)
 
@@ -44,7 +57,12 @@ def llm_eval_section(query: str, section_content: str, llm: LLM) -> bool:
 
 
 def llm_batch_eval_sections(
-    query: str, section_contents: list[str], llm: LLM, use_threads: bool = True
+    query: str,
+    section_contents: list[str],
+    llm: LLM,
+    titles: list[str],
+    metadata_list: list[dict[str, str | list[str]]],
+    use_threads: bool = True,
 ) -> list[bool]:
     if DISABLE_LLM_DOC_RELEVANCE:
         raise RuntimeError(
@@ -54,8 +72,10 @@ def llm_batch_eval_sections(
 
     if use_threads:
         functions_with_args: list[tuple[Callable, tuple]] = [
-            (llm_eval_section, (query, section_content, llm))
+            (llm_eval_section, (query, section_content, llm, title, metadata))
-            for section_content in section_contents
+            for section_content, title, metadata in zip(
+                section_contents, titles, metadata_list
+            )
         ]
 
         logger.debug(
@@ -70,6 +90,8 @@ def llm_batch_eval_sections(
 
     else:
         return [
-            llm_eval_section(query, section_content, llm)
+            llm_eval_section(query, section_content, llm, title, metadata)
-            for section_content in section_contents
+            for section_content, title, metadata in zip(
+                section_contents, titles, metadata_list
+            )
         ]

web/src/components/search/SearchResultsDisplay.tsx

@@ -14,6 +14,7 @@ import { useContext, useEffect, useState } from "react";
 import { Tooltip } from "../tooltip/Tooltip";
 import KeyboardSymbol from "@/lib/browserUtilities";
 import { SettingsContext } from "../settings/SettingsProvider";
+import { DISABLE_LLM_DOC_RELEVANCE } from "@/lib/constants";
 
 const getSelectedDocumentIds = (
   documents: SearchDanswerDocument[],
@@ -140,6 +141,7 @@ export const SearchResultsDisplay = ({
           showAll ||
           (searchResponse &&
             searchResponse.additional_relevance &&
+            searchResponse.additional_relevance[doc.document_id] &&
             searchResponse.additional_relevance[doc.document_id].relevant) ||
           doc.is_relevant
         );
@@ -175,7 +177,8 @@ export const SearchResultsDisplay = ({
         <div className="mt-4">
           <div className="font-bold flex justify-between text-emphasis border-b mb-3 pb-1 border-border text-lg">
             <p>Results</p>
-            {(contentEnriched || searchResponse.additional_relevance) && (
+            {!DISABLE_LLM_DOC_RELEVANCE &&
+              (contentEnriched || searchResponse.additional_relevance) && (
                 <Tooltip delayDuration={1000} content={`${commandSymbol}O`}>
                   <button
                     onClick={() => {
@@ -204,6 +207,7 @@ export const SearchResultsDisplay = ({
                         : agenticResults
                           ? "Focus"
                           : "Show All"}
+
                       <span className="ml-1">
                         {!sweep ? (
                           <BroomIcon className="h-4 w-4" />

web/src/components/search/SearchSection.tsx

@@ -170,11 +170,12 @@ export const SearchSection = ({
     if (existingSearchIdRaw == null) {
       return;
     }
-    function extractFirstUserMessage(
+    function extractFirstMessageByType(
-      chatSession: SearchSession
+      chatSession: SearchSession,
+      messageType: "user" | "assistant"
     ): string | null {
       const userMessage = chatSession?.messages.find(
-        (msg) => msg.message_type === "user"
+        (msg) => msg.message_type === messageType
       );
       return userMessage ? userMessage.message : null;
     }
@@ -184,14 +185,18 @@ export const SearchSection = ({
         `/api/query/search-session/${existingSearchessionId}`
       );
       const searchSession = (await response.json()) as SearchSession;
-      const message = extractFirstUserMessage(searchSession);
+      const userMessage = extractFirstMessageByType(searchSession, "user");
+      const assistantMessage = extractFirstMessageByType(
+        searchSession,
+        "assistant"
+      );
 
-      if (message) {
+      if (userMessage) {
-        setQuery(message);
+        setQuery(userMessage);
         const danswerDocs: SearchResponse = {
           documents: searchSession.documents,
           suggestedSearchType: null,
-          answer: null,
+          answer: assistantMessage || "Search response not found",
           quotes: null,
           selectedDocIndices: null,
           error: null,