Continue Generating (#2286)

* add stop reason

* add initial propagation

* add continue generating full functionality

* proper continue across chat session

* add new look

* propagate proper types

* fix typing

* cleaner continue generating functionality

* update types

* remove unused imports

* proper infodump

* temp

* add standardized stream handling

* validateing chosen tool args

* properly handle tools

* proper ports

* remove logs + build

* minor typing fix

* fix more minor typing issues

* add stashed reversion for tool call chunks

* ignore model dump types

* remove stop stream

* fix typing

backend/danswer/chat/models.py

@@ -1,5 +1,6 @@
 from collections.abc import Iterator
 from datetime import datetime
+from enum import Enum
 from typing import Any
 
 from pydantic import BaseModel
@@ -44,6 +45,20 @@ class QADocsResponse(RetrievalDocs):
         return initial_dict
 
 
+class StreamStopReason(Enum):
+    CONTEXT_LENGTH = "context_length"
+    CANCELLED = "cancelled"
+
+
+class StreamStopInfo(BaseModel):
+    stop_reason: StreamStopReason
+
+    def model_dump(self, *args: list, **kwargs: dict[str, Any]) -> dict[str, Any]:  # type: ignore
+        data = super().model_dump(mode="json", *args, **kwargs)  # type: ignore
+        data["stop_reason"] = self.stop_reason.name
+        return data
+
+
 class LLMRelevanceFilterResponse(BaseModel):
     relevant_chunk_indices: list[int]
 
@@ -144,6 +159,7 @@ AnswerQuestionPossibleReturn = (
     | ImageGenerationDisplay
     | CustomToolResponse
     | StreamingError
+    | StreamStopInfo
 )
 
 

backend/danswer/llm/answering/answer.py

@@ -1,5 +1,6 @@
 from collections.abc import Callable
 from collections.abc import Iterator
+from typing import Any
 from typing import cast
 from uuid import uuid4
 
@@ -12,6 +13,8 @@ from danswer.chat.models import AnswerQuestionPossibleReturn
 from danswer.chat.models import CitationInfo
 from danswer.chat.models import DanswerAnswerPiece
 from danswer.chat.models import LlmDoc
+from danswer.chat.models import StreamStopInfo
+from danswer.chat.models import StreamStopReason
 from danswer.configs.chat_configs import QA_PROMPT_OVERRIDE
 from danswer.file_store.utils import InMemoryChatFile
 from danswer.llm.answering.models import AnswerStyleConfig
@@ -35,7 +38,7 @@ from danswer.llm.answering.stream_processing.quotes_processing import (
 from danswer.llm.answering.stream_processing.utils import DocumentIdOrderMapping
 from danswer.llm.answering.stream_processing.utils import map_document_id_order
 from danswer.llm.interfaces import LLM
-from danswer.llm.utils import message_generator_to_string_generator
+from danswer.llm.interfaces import ToolChoiceOptions
 from danswer.natural_language_processing.utils import get_tokenizer
 from danswer.tools.custom.custom_tool_prompt_builder import (
     build_user_message_for_custom_tool_for_non_tool_calling_llm,
@@ -190,7 +193,9 @@ class Answer:
 
     def _raw_output_for_explicit_tool_calling_llms(
         self,
-    ) -> Iterator[str | ToolCallKickoff | ToolResponse | ToolCallFinalResult]:
+    ) -> Iterator[
+        str | StreamStopInfo | ToolCallKickoff | ToolResponse | ToolCallFinalResult
+    ]:
         prompt_builder = AnswerPromptBuilder(self.message_history, self.llm.config)
 
         tool_call_chunk: AIMessageChunk | None = None
@@ -225,6 +230,7 @@ class Answer:
                     self.tools, self.force_use_tool
                 )
             ]
+
             for message in self.llm.stream(
                 prompt=prompt,
                 tools=final_tool_definitions if final_tool_definitions else None,
@@ -298,21 +304,41 @@ class Answer:
             yield tool_runner.tool_final_result()
 
             prompt = prompt_builder.build(tool_call_summary=tool_call_summary)
-            for token in message_generator_to_string_generator(
+
-                self.llm.stream(
+            yield from self._process_llm_stream(
-                    prompt=prompt,
+                prompt=prompt,
-                    tools=[tool.tool_definition() for tool in self.tools],
+                tools=[tool.tool_definition() for tool in self.tools],
-                )
+            )
-            ):
-                if self.is_cancelled:
-                    return
-                yield token
 
             return
 
+    # This method processes the LLM stream and yields the content or stop information
+    def _process_llm_stream(
+        self,
+        prompt: Any,
+        tools: list[dict] | None = None,
+        tool_choice: ToolChoiceOptions | None = None,
+    ) -> Iterator[str | StreamStopInfo]:
+        for message in self.llm.stream(
+            prompt=prompt, tools=tools, tool_choice=tool_choice
+        ):
+            if isinstance(message, AIMessageChunk):
+                if message.content:
+                    if self.is_cancelled:
+                        return StreamStopInfo(stop_reason=StreamStopReason.CANCELLED)
+                    yield cast(str, message.content)
+
+            if (
+                message.additional_kwargs.get("usage_metadata", {}).get("stop")
+                == "length"
+            ):
+                yield StreamStopInfo(stop_reason=StreamStopReason.CONTEXT_LENGTH)
+
     def _raw_output_for_non_explicit_tool_calling_llms(
         self,
-    ) -> Iterator[str | ToolCallKickoff | ToolResponse | ToolCallFinalResult]:
+    ) -> Iterator[
+        str | StreamStopInfo | ToolCallKickoff | ToolResponse | ToolCallFinalResult
+    ]:
         prompt_builder = AnswerPromptBuilder(self.message_history, self.llm.config)
         chosen_tool_and_args: tuple[Tool, dict] | None = None
 
@@ -387,13 +413,10 @@ class Answer:
                 )
             )
             prompt = prompt_builder.build()
-            for token in message_generator_to_string_generator(
+            yield from self._process_llm_stream(
-                self.llm.stream(prompt=prompt)
+                prompt=prompt,
-            ):
+                tools=None,
-                if self.is_cancelled:
+            )
-                    return
-                yield token
-
             return
 
         tool, tool_args = chosen_tool_and_args
@@ -447,12 +470,8 @@ class Answer:
         yield final
 
         prompt = prompt_builder.build()
-        for token in message_generator_to_string_generator(
+
-            self.llm.stream(prompt=prompt)
+        yield from self._process_llm_stream(prompt=prompt, tools=None)
-        ):
-            if self.is_cancelled:
-                return
-            yield token
 
     @property
     def processed_streamed_output(self) -> AnswerStream:
@@ -470,7 +489,7 @@ class Answer:
         )
 
         def _process_stream(
-            stream: Iterator[ToolCallKickoff | ToolResponse | str],
+            stream: Iterator[ToolCallKickoff | ToolResponse | str | StreamStopInfo],
         ) -> AnswerStream:
             message = None
 
@@ -524,12 +543,15 @@ class Answer:
                     answer_style_configs=self.answer_style_config,
                 )
 
-                def _stream() -> Iterator[str]:
+                def _stream() -> Iterator[str | StreamStopInfo]:
-                    if message:
+                    yield cast(str | StreamStopInfo, message)
-                        yield cast(str, message)
+                    yield from (cast(str | StreamStopInfo, item) for item in stream)
-                    yield from cast(Iterator[str], stream)
 
-                yield from process_answer_stream_fn(_stream())
+                for item in _stream():
+                    if isinstance(item, StreamStopInfo):
+                        yield item
+                    else:
+                        yield from process_answer_stream_fn(iter([item]))
 
         processed_stream = []
         for processed_packet in _process_stream(output_generator):

backend/danswer/llm/answering/stream_processing/citation_processing.py

@@ -11,7 +11,6 @@ from danswer.llm.answering.stream_processing.utils import DocumentIdOrderMapping
 from danswer.prompts.constants import TRIPLE_BACKTICK
 from danswer.utils.logger import setup_logger
 
-
 logger = setup_logger()
 
 
@@ -204,7 +203,9 @@ def extract_citations_from_stream(
 def build_citation_processor(
     context_docs: list[LlmDoc], doc_id_to_rank_map: DocumentIdOrderMapping
 ) -> StreamProcessor:
-    def stream_processor(tokens: Iterator[str]) -> AnswerQuestionStreamReturn:
+    def stream_processor(
+        tokens: Iterator[str],
+    ) -> AnswerQuestionStreamReturn:
         yield from extract_citations_from_stream(
             tokens=tokens,
             context_docs=context_docs,

backend/danswer/llm/answering/stream_processing/quotes_processing.py

@@ -285,7 +285,9 @@ def process_model_tokens(
 def build_quotes_processor(
     context_docs: list[LlmDoc], is_json_prompt: bool
 ) -> Callable[[Iterator[str]], AnswerQuestionStreamReturn]:
-    def stream_processor(tokens: Iterator[str]) -> AnswerQuestionStreamReturn:
+    def stream_processor(
+        tokens: Iterator[str],
+    ) -> AnswerQuestionStreamReturn:
         yield from process_model_tokens(
             tokens=tokens,
             context_docs=context_docs,

backend/danswer/llm/chat_llm.py

@@ -138,7 +138,9 @@ def _convert_message_to_dict(message: BaseMessage) -> dict:
 
 
 def _convert_delta_to_message_chunk(
-    _dict: dict[str, Any], curr_msg: BaseMessage | None
+    _dict: dict[str, Any],
+    curr_msg: BaseMessage | None,
+    stop_reason: str | None = None,
 ) -> BaseMessageChunk:
     """Adapted from langchain_community.chat_models.litellm._convert_delta_to_message_chunk"""
     role = _dict.get("role") or (_base_msg_to_role(curr_msg) if curr_msg else None)
@@ -163,12 +165,23 @@ def _convert_delta_to_message_chunk(
                 args=tool_call.function.arguments,
                 index=0,  # only support a single tool call atm
             )
+
             return AIMessageChunk(
                 content=content,
-                additional_kwargs=additional_kwargs,
                 tool_call_chunks=[tool_call_chunk],
+                additional_kwargs={
+                    "usage_metadata": {"stop": stop_reason},
+                    **additional_kwargs,
+                },
             )
-        return AIMessageChunk(content=content, additional_kwargs=additional_kwargs)
+
+        return AIMessageChunk(
+            content=content,
+            additional_kwargs={
+                "usage_metadata": {"stop": stop_reason},
+                **additional_kwargs,
+            },
+        )
     elif role == "system":
         return SystemMessageChunk(content=content)
     elif role == "function":
@@ -206,7 +219,7 @@ class DefaultMultiLLM(LLM):
         self._api_version = api_version
         self._custom_llm_provider = custom_llm_provider
 
-        # This can be used to store the maximum output tkoens for this model.
+        # This can be used to store the maximum output tokens for this model.
         # self._max_output_tokens = (
         #     max_output_tokens
         #     if max_output_tokens is not None
@@ -349,10 +362,16 @@ class DefaultMultiLLM(LLM):
         )
         try:
             for part in response:
-                if len(part["choices"]) == 0:
+                if not part["choices"]:
                     continue
-                delta = part["choices"][0]["delta"]
+
-                message_chunk = _convert_delta_to_message_chunk(delta, output)
+                choice = part["choices"][0]
+                message_chunk = _convert_delta_to_message_chunk(
+                    choice["delta"],
+                    output,
+                    stop_reason=choice["finish_reason"],
+                )
+
                 if output is None:
                     output = message_chunk
                 else:

web/src/app/chat/ChatPage.tsx

@@ -65,7 +65,12 @@ import { FiArrowDown } from "react-icons/fi";
 import { ChatIntro } from "./ChatIntro";
 import { AIMessage, HumanMessage } from "./message/Messages";
 import { StarterMessage } from "./StarterMessage";
-import { AnswerPiecePacket, DanswerDocument } from "@/lib/search/interfaces";
+import {
+  AnswerPiecePacket,
+  DanswerDocument,
+  StreamStopInfo,
+  StreamStopReason,
+} from "@/lib/search/interfaces";
 import { buildFilters } from "@/lib/search/utils";
 import { SettingsContext } from "@/components/settings/SettingsProvider";
 import Dropzone from "react-dropzone";
@@ -94,6 +99,7 @@ import ExceptionTraceModal from "@/components/modals/ExceptionTraceModal";
 
 import { SEARCH_TOOL_NAME } from "./tools/constants";
 import { useUser } from "@/components/user/UserProvider";
+import { Stop } from "@phosphor-icons/react";
 
 const TEMP_USER_MESSAGE_ID = -1;
 const TEMP_ASSISTANT_MESSAGE_ID = -2;
@@ -338,6 +344,7 @@ export function ChatPage({
         }
         return;
       }
+
       clearSelectedDocuments();
       setIsFetchingChatMessages(true);
       const response = await fetch(
@@ -624,6 +631,24 @@ export function ChatPage({
   const currentRegenerationState = (): RegenerationState | null => {
     return regenerationState.get(currentSessionId()) || null;
   };
+  const [canContinue, setCanContinue] = useState<Map<number | null, boolean>>(
+    new Map([[null, false]])
+  );
+
+  const updateCanContinue = (newState: boolean, sessionId?: number | null) => {
+    setCanContinue((prevState) => {
+      const newCanContinueState = new Map(prevState);
+      newCanContinueState.set(
+        sessionId !== undefined ? sessionId : currentSessionId(),
+        newState
+      );
+      return newCanContinueState;
+    });
+  };
+
+  const currentCanContinue = (): boolean => {
+    return canContinue.get(currentSessionId()) || false;
+  };
 
   const currentSessionChatState = currentChatState();
   const currentSessionRegenerationState = currentRegenerationState();
@@ -864,6 +889,13 @@ export function ChatPage({
     }
   };
 
+  const continueGenerating = () => {
+    onSubmit({
+      messageOverride:
+        "Continue Generating (pick up exactly where you left off)",
+    });
+  };
+
   const onSubmit = async ({
     messageIdToResend,
     messageOverride,
@@ -884,6 +916,7 @@ export function ChatPage({
     regenerationRequest?: RegenerationRequest | null;
   } = {}) => {
     let frozenSessionId = currentSessionId();
+    updateCanContinue(false, frozenSessionId);
 
     if (currentChatState() != "input") {
       setPopup({
@@ -978,6 +1011,8 @@ export function ChatPage({
     let messageUpdates: Message[] | null = null;
 
     let answer = "";
+
+    let stopReason: StreamStopReason | null = null;
     let query: string | null = null;
     let retrievalType: RetrievalType =
       selectedDocuments.length > 0
@@ -1174,6 +1209,11 @@ export function ChatPage({
               stackTrace = (packet as StreamingError).stack_trace;
             } else if (Object.hasOwn(packet, "message_id")) {
               finalMessage = packet as BackendMessage;
+            } else if (Object.hasOwn(packet, "stop_reason")) {
+              const stop_reason = (packet as StreamStopInfo).stop_reason;
+              if (stop_reason === StreamStopReason.CONTEXT_LENGTH) {
+                updateCanContinue(true, frozenSessionId);
+              }
             }
 
             // on initial message send, we insert a dummy system message
@@ -1237,6 +1277,7 @@ export function ChatPage({
                 alternateAssistantID: alternativeAssistant?.id,
                 stackTrace: stackTrace,
                 overridden_model: finalMessage?.overridden_model,
+                stopReason: stopReason,
               },
             ]);
           }
@@ -1835,6 +1876,12 @@ export function ChatPage({
                                     }
                                   >
                                     <AIMessage
+                                      continueGenerating={
+                                        i == messageHistory.length - 1 &&
+                                        currentCanContinue()
+                                          ? continueGenerating
+                                          : undefined
+                                      }
                                       overriddenModel={message.overridden_model}
                                       regenerate={createRegenerator({
                                         messageId: message.messageId,

web/src/app/chat/interfaces.ts

@@ -2,6 +2,7 @@ import {
   DanswerDocument,
   Filters,
   SearchDanswerDocument,
+  StreamStopReason,
 } from "@/lib/search/interfaces";
 
 export enum RetrievalType {
@@ -89,6 +90,7 @@ export interface Message {
   alternateAssistantID?: number | null;
   stackTrace?: string | null;
   overridden_model?: string;
+  stopReason?: StreamStopReason | null;
 }
 
 export interface BackendChatSession {

web/src/app/chat/lib.tsx

@@ -2,6 +2,7 @@ import {
   AnswerPiecePacket,
   DanswerDocument,
   Filters,
+  StreamStopInfo,
 } from "@/lib/search/interfaces";
 import { handleSSEStream, handleStream } from "@/lib/search/streamingUtils";
 import { ChatState, FeedbackType } from "./types";
@@ -111,7 +112,8 @@ export type PacketType =
   | DocumentsResponse
   | ImageGenerationDisplay
   | StreamingError
-  | MessageResponseIDInfo;
+  | MessageResponseIDInfo
+  | StreamStopInfo;
 
 export async function* sendMessage({
   regenerate,

web/src/app/chat/message/ContinueMessage.tsx

@@ -0,0 +1,37 @@
+import { EmphasizedClickable } from "@/components/BasicClickable";
+import { useEffect, useState } from "react";
+import { FiBook, FiPlayCircle } from "react-icons/fi";
+
+export function ContinueGenerating({
+  handleContinueGenerating,
+}: {
+  handleContinueGenerating: () => void;
+}) {
+  const [showExplanation, setShowExplanation] = useState(false);
+
+  useEffect(() => {
+    const timer = setTimeout(() => {
+      setShowExplanation(true);
+    }, 1000);
+
+    return () => clearTimeout(timer);
+  }, []);
+
+  return (
+    <div className="flex justify-center w-full">
+      <div className="relative group">
+        <EmphasizedClickable onClick={handleContinueGenerating}>
+          <>
+            <FiPlayCircle className="mr-2" />
+            Continue Generation
+          </>
+        </EmphasizedClickable>
+        {showExplanation && (
+          <div className="absolute bottom-full left-1/2 transform -translate-x-1/2 mb-2 px-3 py-1 bg-gray-800 text-white text-xs rounded-lg opacity-0 group-hover:opacity-100 transition-opacity duration-300 whitespace-nowrap">
+            LLM reached its token limit. Click to continue.
+          </div>
+        )}
+      </div>
+    </div>
+  );
+}

web/src/app/chat/message/Messages.tsx

@@ -65,6 +65,8 @@ import GeneratingImageDisplay from "../tools/GeneratingImageDisplay";
 import RegenerateOption from "../RegenerateOption";
 import { LlmOverride } from "@/lib/hooks";
 import ExceptionTraceModal from "@/components/modals/ExceptionTraceModal";
+import { EmphasizedClickable } from "@/components/BasicClickable";
+import { ContinueGenerating } from "./ContinueMessage";
 
 const TOOLS_WITH_CUSTOM_HANDLING = [
   SEARCH_TOOL_NAME,
@@ -123,6 +125,7 @@ function FileDisplay({
 export const AIMessage = ({
   regenerate,
   overriddenModel,
+  continueGenerating,
   shared,
   isActive,
   toggleDocumentSelection,
@@ -150,6 +153,7 @@ export const AIMessage = ({
 }: {
   shared?: boolean;
   isActive?: boolean;
+  continueGenerating?: () => void;
   otherMessagesCanSwitchTo?: number[];
   onMessageSelection?: (messageId: number) => void;
   selectedDocuments?: DanswerDocument[] | null;
@@ -283,11 +287,12 @@ export const AIMessage = ({
               size="small"
               assistant={alternativeAssistant || currentPersona}
             />
+
             <div className="w-full">
               <div className="max-w-message-max break-words">
                 <div className="w-full ml-4">
                   <div className="max-w-message-max break-words">
-                    {(!toolCall || toolCall.tool_name === SEARCH_TOOL_NAME) && (
+                    {!toolCall || toolCall.tool_name === SEARCH_TOOL_NAME ? (
                       <>
                         {query !== undefined &&
                           handleShowRetrieved !== undefined &&
@@ -315,7 +320,8 @@ export const AIMessage = ({
                             </div>
                           )}
                       </>
-                    )}
+                    ) : null}
+
                     {toolCall &&
                       !TOOLS_WITH_CUSTOM_HANDLING.includes(
                         toolCall.tool_name
@@ -633,6 +639,11 @@ export const AIMessage = ({
             </div>
           </div>
         </div>
+        {(!toolCall || toolCall.tool_name === SEARCH_TOOL_NAME) &&
+          !query &&
+          continueGenerating && (
+            <ContinueGenerating handleContinueGenerating={continueGenerating} />
+          )}
       </div>
     </div>
   );

web/src/app/chat/message/SkippedSearch.tsx

@@ -27,7 +27,7 @@ export function SkippedSearch({
   handleForceSearch: () => void;
 }) {
   return (
-    <div className="flex  text-sm !pt-0 p-1">
+    <div className="flex text-sm !pt-0 p-1">
       <div className="flex mb-auto">
         <FiBook className="my-auto flex-none mr-2" size={14} />
         <div className="my-auto cursor-default">

web/src/lib/search/interfaces.ts

@@ -19,6 +19,15 @@ export interface AnswerPiecePacket {
   answer_piece: string;
 }
 
+export enum StreamStopReason {
+  CONTEXT_LENGTH = "CONTEXT_LENGTH",
+  CANCELLED = "CANCELLED",
+}
+
+export interface StreamStopInfo {
+  stop_reason: StreamStopReason;
+}
+
 export interface ErrorMessagePacket {
   error: string;
 }