Touchup for Multilingual Users (#1725)

backend/alembic/versions/bc9771dccadf_create_usage_reports_table.py

@@ -12,8 +12,8 @@ import fastapi_users_db_sqlalchemy
 # revision identifiers, used by Alembic.
 revision = "bc9771dccadf"
 down_revision = "0568ccf46a6b"
-branch_labels = None
+branch_labels: None = None
-depends_on = None
+depends_on: None = None
 
 
 def upgrade() -> None:

backend/danswer/configs/chat_configs.py

@@ -64,6 +64,14 @@ TITLE_CONTENT_RATIO = max(
 # A list of languages passed to the LLM to rephase the query
 # For example "English,French,Spanish", be sure to use the "," separator
 MULTILINGUAL_QUERY_EXPANSION = os.environ.get("MULTILINGUAL_QUERY_EXPANSION") or None
+LANGUAGE_HINT = "\n" + (
+    os.environ.get("LANGUAGE_HINT")
+    or "IMPORTANT: Respond in the same language as my query!"
+)
+LANGUAGE_CHAT_NAMING_HINT = (
+    os.environ.get("LANGUAGE_CHAT_NAMING_HINT")
+    or "The name of the conversation must be in the same language as the user query."
+)
 
 # Stops streaming answers back to the UI if this pattern is seen:
 STOP_STREAM_PAT = os.environ.get("STOP_STREAM_PAT") or None

backend/danswer/llm/answering/prompts/quotes_prompt.py

@@ -1,13 +1,13 @@
 from langchain.schema.messages import HumanMessage
 
 from danswer.chat.models import LlmDoc
+from danswer.configs.chat_configs import LANGUAGE_HINT
 from danswer.configs.chat_configs import MULTILINGUAL_QUERY_EXPANSION
 from danswer.configs.chat_configs import QA_PROMPT_OVERRIDE
 from danswer.llm.answering.models import PromptConfig
 from danswer.prompts.direct_qa_prompts import CONTEXT_BLOCK
 from danswer.prompts.direct_qa_prompts import HISTORY_BLOCK
 from danswer.prompts.direct_qa_prompts import JSON_PROMPT
-from danswer.prompts.direct_qa_prompts import LANGUAGE_HINT
 from danswer.prompts.direct_qa_prompts import WEAK_LLM_PROMPT
 from danswer.prompts.prompt_utils import add_date_time_to_prompt
 from danswer.prompts.prompt_utils import build_complete_context_str

backend/danswer/prompts/chat_prompts.py

@@ -188,7 +188,7 @@ Query:
 
 
 CHAT_NAMING = f"""
-Given the following conversation, provide a SHORT name for the conversation.
+Given the following conversation, provide a SHORT name for the conversation.{{language_hint_or_empty}}
 IMPORTANT: TRY NOT TO USE MORE THAN 5 WORDS, MAKE IT AS CONCISE AS POSSIBLE.
 Focus the name on the important keywords to convey the topic of the conversation.
 

backend/danswer/prompts/direct_qa_prompts.py

@@ -41,12 +41,6 @@ Hint: Make the answer as DETAILED as possible and respond in JSON format! \
 Quotes MUST be EXACT substrings from provided documents!
 """.strip()
 
-
-LANGUAGE_HINT = """
-IMPORTANT: Respond in the same language as my query!
-"""
-
-
 CONTEXT_BLOCK = f"""
 REFERENCE DOCUMENTS:
 {GENERAL_SEP_PAT}

backend/danswer/prompts/miscellaneous_prompts.py

@@ -2,9 +2,7 @@
 
 LANGUAGE_REPHRASE_PROMPT = """
 Translate query to {target_language}.
-If the query at the end is already in {target_language}, \
+If the query at the end is already in {target_language}, simply repeat the ORIGINAL query back to me, EXACTLY as is with no edits.
-simply repeat the ORIGINAL query back to me, EXACTLY as is with no edits.
-
 If the query below is not in {target_language}, translate it into {target_language}.
 
 Query:

backend/danswer/prompts/prompt_utils.py

@@ -5,6 +5,7 @@ from typing import cast
 from langchain_core.messages import BaseMessage
 
 from danswer.chat.models import LlmDoc
+from danswer.configs.chat_configs import LANGUAGE_HINT
 from danswer.configs.chat_configs import MULTILINGUAL_QUERY_EXPANSION
 from danswer.configs.constants import DocumentSource
 from danswer.db.models import Prompt
@@ -12,7 +13,6 @@ from danswer.llm.answering.models import PromptConfig
 from danswer.prompts.chat_prompts import ADDITIONAL_INFO
 from danswer.prompts.chat_prompts import CITATION_REMINDER
 from danswer.prompts.constants import CODE_BLOCK_PAT
-from danswer.prompts.direct_qa_prompts import LANGUAGE_HINT
 from danswer.search.models import InferenceChunk
 
 

backend/danswer/prompts/token_counts.py

@@ -1,14 +1,13 @@
+from danswer.configs.chat_configs import LANGUAGE_HINT
 from danswer.llm.utils import check_number_of_tokens
 from danswer.prompts.chat_prompts import ADDITIONAL_INFO
 from danswer.prompts.chat_prompts import CHAT_USER_PROMPT
 from danswer.prompts.chat_prompts import CITATION_REMINDER
 from danswer.prompts.chat_prompts import REQUIRE_CITATION_STATEMENT
 from danswer.prompts.constants import DEFAULT_IGNORE_STATEMENT
-from danswer.prompts.direct_qa_prompts import LANGUAGE_HINT
 from danswer.prompts.prompt_utils import get_current_llm_day_time
 
-# tokens outside of the actual persona's "user_prompt" that make up the end
+# tokens outside of the actual persona's "user_prompt" that make up the end user message
-# user message
 CHAT_USER_PROMPT_WITH_CONTEXT_OVERHEAD_TOKEN_CNT = check_number_of_tokens(
     CHAT_USER_PROMPT.format(
         context_docs_str="",

backend/danswer/secondary_llm_flows/chat_session_naming.py

@@ -1,4 +1,6 @@
 from danswer.chat.chat_utils import combine_message_chain
+from danswer.configs.chat_configs import LANGUAGE_CHAT_NAMING_HINT
+from danswer.configs.chat_configs import MULTILINGUAL_QUERY_EXPANSION
 from danswer.configs.model_configs import GEN_AI_HISTORY_CUTOFF
 from danswer.db.models import ChatMessage
 from danswer.llm.interfaces import LLM
@@ -18,10 +20,18 @@ def get_renamed_conversation_name(
         messages=full_history, token_limit=GEN_AI_HISTORY_CUTOFF
     )
 
+    language_hint = (
+        f"\n{LANGUAGE_CHAT_NAMING_HINT.strip()}"
+        if bool(MULTILINGUAL_QUERY_EXPANSION)
+        else ""
+    )
+
     prompt_msgs = [
         {
             "role": "user",
-            "content": CHAT_NAMING.format(chat_history=history_str),
+            "content": CHAT_NAMING.format(
+                language_hint_or_empty=language_hint, chat_history=history_str
+            ),
         },
     ]
 

deployment/docker_compose/docker-compose.dev.yml

@@ -63,6 +63,8 @@ services:
       - HYBRID_ALPHA=${HYBRID_ALPHA:-}  # Hybrid Search Alpha (0 for entirely keyword, 1 for entirely vector)
       - EDIT_KEYWORD_QUERY=${EDIT_KEYWORD_QUERY:-}
       - MULTILINGUAL_QUERY_EXPANSION=${MULTILINGUAL_QUERY_EXPANSION:-}
+      - LANGUAGE_HINT=${LANGUAGE_HINT:-}
+      - LANGUAGE_CHAT_NAMING_HINT=${LANGUAGE_CHAT_NAMING_HINT:-}
       - QA_PROMPT_OVERRIDE=${QA_PROMPT_OVERRIDE:-}
       # Other services
       - POSTGRES_HOST=relational_db
@@ -140,6 +142,8 @@ services:
       - HYBRID_ALPHA=${HYBRID_ALPHA:-}  # Hybrid Search Alpha (0 for entirely keyword, 1 for entirely vector)
       - EDIT_KEYWORD_QUERY=${EDIT_KEYWORD_QUERY:-}
       - MULTILINGUAL_QUERY_EXPANSION=${MULTILINGUAL_QUERY_EXPANSION:-}
+      - LANGUAGE_HINT=${LANGUAGE_HINT:-}
+      - LANGUAGE_CHAT_NAMING_HINT=${LANGUAGE_CHAT_NAMING_HINT:-}
       - QA_PROMPT_OVERRIDE=${QA_PROMPT_OVERRIDE:-}
       # Other Services
       - POSTGRES_HOST=relational_db

deployment/docker_compose/docker-compose.gpu-dev.yml

@@ -59,6 +59,8 @@ services:
       - HYBRID_ALPHA=${HYBRID_ALPHA:-}  # Hybrid Search Alpha (0 for entirely keyword, 1 for entirely vector)
       - EDIT_KEYWORD_QUERY=${EDIT_KEYWORD_QUERY:-}
       - MULTILINGUAL_QUERY_EXPANSION=${MULTILINGUAL_QUERY_EXPANSION:-}
+      - LANGUAGE_HINT=${LANGUAGE_HINT:-}
+      - LANGUAGE_CHAT_NAMING_HINT=${LANGUAGE_CHAT_NAMING_HINT:-}
       - QA_PROMPT_OVERRIDE=${QA_PROMPT_OVERRIDE:-}
       # Other services
       - POSTGRES_HOST=relational_db
@@ -132,6 +134,8 @@ services:
       - HYBRID_ALPHA=${HYBRID_ALPHA:-}  # Hybrid Search Alpha (0 for entirely keyword, 1 for entirely vector)
       - EDIT_KEYWORD_QUERY=${EDIT_KEYWORD_QUERY:-}
       - MULTILINGUAL_QUERY_EXPANSION=${MULTILINGUAL_QUERY_EXPANSION:-}
+      - LANGUAGE_HINT=${LANGUAGE_HINT:-}
+      - LANGUAGE_CHAT_NAMING_HINT=${LANGUAGE_CHAT_NAMING_HINT:-}
       - QA_PROMPT_OVERRIDE=${QA_PROMPT_OVERRIDE:-}
       # Other Services
       - POSTGRES_HOST=relational_db

deployment/docker_compose/env.multilingual.template

@@ -6,6 +6,9 @@
 
 # Rephrase the user query in specified languages using LLM, use comma separated values
 MULTILINGUAL_QUERY_EXPANSION="English, French"
+# Change the below to suit your specific needs, can be more explicit about the language of the response
+LANGUAGE_HINT="IMPORTANT: Respond in the same language as my query!"
+LANGUAGE_CHAT_NAMING_HINT="The name of the conversation must be in the same language as the user query."
 
 # A recent MIT license multilingual model: https://huggingface.co/intfloat/multilingual-e5-small
 DOCUMENT_ENCODER_MODEL="intfloat/multilingual-e5-small"

deployment/helm/values.yaml

@@ -411,6 +411,8 @@ configMap:
   HYBRID_ALPHA: ""
   EDIT_KEYWORD_QUERY: ""
   MULTILINGUAL_QUERY_EXPANSION: ""
+  LANGUAGE_HINT: ""
+  LANGUAGE_CHAT_NAMING_HINT: ""
   QA_PROMPT_OVERRIDE: ""
   # Internet Search Tool
   BING_API_KEY: ""

deployment/kubernetes/env-configmap.yaml

@@ -33,6 +33,8 @@ data:
   HYBRID_ALPHA: ""
   EDIT_KEYWORD_QUERY: ""
   MULTILINGUAL_QUERY_EXPANSION: ""
+  LANGUAGE_HINT: ""
+  LANGUAGE_CHAT_NAMING_HINT: ""
   QA_PROMPT_OVERRIDE: ""
   # Other Services
   POSTGRES_HOST: "relational-db-service"