Fixes #911: handle prompt too long errors

packages/shared/src/env.server.ts

@@ -231,6 +231,8 @@ export const env = createEnv({
 
         SOURCEBOT_CHAT_MODEL_TEMPERATURE: numberSchema.default(0.3),
         SOURCEBOT_CHAT_MAX_STEP_COUNT: numberSchema.default(20),
+        SOURCEBOT_CHAT_FILE_MAX_CHARACTERS: numberSchema.default(100_000),
+        SOURCEBOT_CHAT_MAX_MESSAGE_HISTORY: numberSchema.default(50),
 
         DEBUG_WRITE_CHAT_MESSAGES_TO_FILE: booleanSchema.default('false'),
 

packages/web/src/app/api/(server)/chat/route.ts

@@ -2,7 +2,7 @@ import { sew } from "@/actions";
 import { _getConfiguredLanguageModelsFull, _getAISDKLanguageModelAndOptions, _updateChatMessages, _isOwnerOfChat } from "@/features/chat/actions";
 import { createAgentStream } from "@/features/chat/agent";
 import { additionalChatRequestParamsSchema, LanguageModelInfo, SBChatMessage, SearchScope } from "@/features/chat/types";
-import { getAnswerPartFromAssistantMessage, getLanguageModelKey } from "@/features/chat/utils";
+import { getAnswerPartFromAssistantMessage, getLanguageModelKey, isContextWindowError, CONTEXT_WINDOW_USER_MESSAGE } from "@/features/chat/utils";
 import { apiHandler } from "@/lib/apiHandler";
 import { ErrorCode } from "@/lib/errorCodes";
 import { notFound, requestBodySchemaValidationError, ServiceError, serviceErrorResponse } from "@/lib/serviceError";
@@ -11,7 +11,7 @@ import { withOptionalAuthV2 } from "@/withAuthV2";
 import { LanguageModelV2 as AISDKLanguageModelV2 } from "@ai-sdk/provider";
 import * as Sentry from "@sentry/nextjs";
 import { PrismaClient } from "@sourcebot/db";
-import { createLogger } from "@sourcebot/shared";
+import { createLogger, env } from "@sourcebot/shared";
 import { captureEvent } from "@/lib/posthog";
 import {
     createUIMessageStream,
@@ -114,15 +114,17 @@ export const POST = apiHandler(async (req: NextRequest) => {
                         return 'unknown error';
                     }
 
-                    if (typeof error === 'string') {
-                        return error;
-                    }
+                    const errorMessage = (() => {
+                        if (typeof error === 'string') return error;
+                        if (error instanceof Error) return error.message;
+                        return JSON.stringify(error);
+                    })();
 
-                    if (error instanceof Error) {
-                        return error.message;
+                    if (isContextWindowError(errorMessage)) {
+                        return CONTEXT_WINDOW_USER_MESSAGE;
                     }
 
-                    return JSON.stringify(error);
+                    return errorMessage;
                 }
             });
 
@@ -203,6 +205,11 @@ export const createMessageStream = async ({
             }
         }).filter(message => message !== undefined);
 
+    const maxMessages = env.SOURCEBOT_CHAT_MAX_MESSAGE_HISTORY;
+    const trimmedMessageHistory = messageHistory.length > maxMessages
+        ? messageHistory.slice(-maxMessages)
+        : messageHistory;
+
     const stream = createUIMessageStream<SBChatMessage>({
         execute: async ({ writer }) => {
             writer.write({
@@ -238,7 +245,7 @@ export const createMessageStream = async ({
             const researchStream = await createAgentStream({
                 model,
                 providerOptions: modelProviderOptions,
-                inputMessages: messageHistory,
+                inputMessages: trimmedMessageHistory,
                 inputSources: sources,
                 selectedRepos: expandedRepos,
                 onWriteSource: (source) => {

packages/web/src/features/chat/agent.ts

@@ -7,7 +7,7 @@ import { LanguageModel, ModelMessage, StopCondition, streamText } from "ai";
 import { ANSWER_TAG, FILE_REFERENCE_PREFIX, toolNames } from "./constants";
 import { createCodeSearchTool, findSymbolDefinitionsTool, findSymbolReferencesTool, listReposTool, listCommitsTool, readFilesTool } from "./tools";
 import { Source } from "./types";
-import { addLineNumbers, fileReferenceToString } from "./utils";
+import { addLineNumbers, fileReferenceToString, truncateFileContent } from "./utils";
 import _dedent from "dedent";
 
 const dedent = _dedent.withOptions({ alignValues: true });
@@ -60,9 +60,20 @@ export const createAgentStream = async ({
         }))
     ).filter((source) => source !== undefined);
 
+    const maxChars = env.SOURCEBOT_CHAT_FILE_MAX_CHARACTERS;
+    let anyFileTruncated = false;
+    const truncatedFileSources = resolvedFileSources.map((file) => {
+        const { content, wasTruncated } = truncateFileContent(file.source, maxChars);
+        if (wasTruncated) {
+            anyFileTruncated = true;
+        }
+        return { ...file, source: content };
+    });
+
     const systemPrompt = createPrompt({
         repos: selectedRepos,
-        files: resolvedFileSources,
+        files: truncatedFileSources,
+        filesWereTruncated: anyFileTruncated,
     });
 
     const stream = streamText({
@@ -148,6 +159,7 @@ export const createAgentStream = async ({
 const createPrompt = ({
     files,
     repos,
+    filesWereTruncated,
 }: {
     files?: {
         path: string;
@@ -157,6 +169,7 @@ const createPrompt = ({
         revision: string;
     }[],
     repos: string[],
+    filesWereTruncated?: boolean,
 }) => {
     return dedent`
     You are a powerful agentic AI code assistant built into Sourcebot, the world's best code-intelligence platform. Your job is to help developers understand and navigate their large codebases.
@@ -189,6 +202,7 @@ const createPrompt = ({
 
     ${(files && files.length > 0) ? dedent`
         <files>
+        ${filesWereTruncated ? `**Note:** Some files were truncated because they exceeded the character limit. Use the readFiles tool to retrieve specific sections if needed.` : ''}
         The user has mentioned the following files, which are automatically included for analysis.
 
         ${files?.map(file => `<file path="${file.path}" repository="${file.repo}" language="${file.language}" revision="${file.revision}">

packages/web/src/features/chat/components/chatThread/errorBanner.tsx

@@ -2,6 +2,7 @@
 
 import { Button } from '@/components/ui/button';
 import { serviceErrorSchema } from '@/lib/serviceError';
+import { CONTEXT_WINDOW_USER_MESSAGE } from '@/features/chat/utils';
 import { AlertCircle, X } from "lucide-react";
 import { useMemo } from 'react';
 
@@ -33,7 +34,7 @@ export const ErrorBanner = ({ error, isVisible, onClose }: ErrorBannerProps) =>
                     <div className="flex items-center gap-2">
                         <AlertCircle className="h-4 w-4 text-red-600 dark:text-red-400" />
                         <span className="text-sm font-medium text-red-800 dark:text-red-200">
-                            Error occurred
+                            {errorMessage === CONTEXT_WINDOW_USER_MESSAGE ? 'Context limit exceeded' : 'Error occurred'}
                         </span>
                         <span className="text-sm text-red-600 dark:text-red-400">
                             {errorMessage}

packages/web/src/features/chat/tools.ts

@@ -4,7 +4,8 @@ import { InferToolInput, InferToolOutput, InferUITool, tool, ToolUIPart } from "
 import { isServiceError } from "@/lib/utils";
 import { FileSourceResponse, getFileSource, listCommits } from '@/features/git';
 import { findSearchBasedSymbolDefinitions, findSearchBasedSymbolReferences } from "../codeNav/api";
-import { addLineNumbers } from "./utils";
+import { addLineNumbers, truncateFileContent } from "./utils";
+import { env } from "@sourcebot/shared";
 import { toolNames } from "./constants";
 import { listReposQueryParamsSchema } from "@/lib/schemas";
 import { ListReposQueryParams } from "@/lib/types";
@@ -123,13 +124,16 @@ export const readFilesTool = tool({
             return firstError!;
         }
 
-        return (responses as FileSourceResponse[]).map((response) => ({
-            path: response.path,
-            repository: response.repo,
-            language: response.language,
-            source: addLineNumbers(response.source),
-            revision,
-        }));
+        return (responses as FileSourceResponse[]).map((response) => {
+            const { content } = truncateFileContent(response.source, env.SOURCEBOT_CHAT_FILE_MAX_CHARACTERS);
+            return {
+                path: response.path,
+                repository: response.repo,
+                language: response.language,
+                source: addLineNumbers(content),
+                revision,
+            };
+        });
     }
 });
 

packages/web/src/features/chat/utils.test.ts

@@ -1,5 +1,5 @@
 import { expect, test, vi } from 'vitest'
-import { fileReferenceToString, getAnswerPartFromAssistantMessage, groupMessageIntoSteps, repairReferences } from './utils'
+import { fileReferenceToString, getAnswerPartFromAssistantMessage, groupMessageIntoSteps, repairReferences, truncateFileContent, isContextWindowError, CONTEXT_WINDOW_USER_MESSAGE } from './utils'
 import { FILE_REFERENCE_REGEX, ANSWER_TAG } from './constants';
 import { SBChatMessage, SBChatMessagePart } from './types';
 
@@ -351,3 +351,73 @@ test('repairReferences handles malformed inline code blocks', () => {
     const expected = 'See @file:{github.com/sourcebot-dev/sourcebot::packages/web/src/auth.ts} for details.';
     expect(repairReferences(input)).toBe(expected);
 });
+
+// truncateFileContent tests
+
+test('truncateFileContent returns content unchanged when under limit', () => {
+    const source = 'line 1\nline 2\nline 3';
+    const result = truncateFileContent(source, 100);
+    expect(result.content).toBe(source);
+    expect(result.wasTruncated).toBe(false);
+});
+
+test('truncateFileContent returns content unchanged when exactly at limit', () => {
+    const source = 'abcde';
+    const result = truncateFileContent(source, 5);
+    expect(result.content).toBe(source);
+    expect(result.wasTruncated).toBe(false);
+});
+
+test('truncateFileContent truncates at line boundary when over limit', () => {
+    const source = 'line 1\nline 2\nline 3\nline 4\nline 5';
+    // Limit of 20 characters: "line 1\nline 2\nline 3" is 20 chars
+    const result = truncateFileContent(source, 15);
+    expect(result.wasTruncated).toBe(true);
+    expect(result.content).toContain('line 1\nline 2');
+    expect(result.content).toContain('... [truncated:');
+    expect(result.content).not.toContain('line 4');
+});
+
+test('truncateFileContent includes line count information', () => {
+    const source = 'a\nb\nc\nd\ne';
+    const result = truncateFileContent(source, 3);
+    expect(result.wasTruncated).toBe(true);
+    expect(result.content).toMatch(/showing \d+ of 5 lines/);
+});
+
+// isContextWindowError tests
+
+test('isContextWindowError detects OpenAI context length error', () => {
+    expect(isContextWindowError('This model\'s maximum context length is 128000 tokens')).toBe(true);
+});
+
+test('isContextWindowError detects Anthropic prompt too long error', () => {
+    expect(isContextWindowError('prompt is too long: 150000 tokens > 100000 maximum')).toBe(true);
+});
+
+test('isContextWindowError detects context_length_exceeded error', () => {
+    expect(isContextWindowError('context_length_exceeded')).toBe(true);
+});
+
+test('isContextWindowError detects token limit error', () => {
+    expect(isContextWindowError('Request exceeds the maximum number of tokens')).toBe(true);
+});
+
+test('isContextWindowError detects reduce the length error', () => {
+    expect(isContextWindowError('Please reduce the length of the messages')).toBe(true);
+});
+
+test('isContextWindowError detects request too large error', () => {
+    expect(isContextWindowError('request too large')).toBe(true);
+});
+
+test('isContextWindowError returns false for unrelated errors', () => {
+    expect(isContextWindowError('Internal server error')).toBe(false);
+    expect(isContextWindowError('Rate limit exceeded')).toBe(false);
+    expect(isContextWindowError('Invalid API key')).toBe(false);
+});
+
+test('CONTEXT_WINDOW_USER_MESSAGE is a non-empty string', () => {
+    expect(typeof CONTEXT_WINDOW_USER_MESSAGE).toBe('string');
+    expect(CONTEXT_WINDOW_USER_MESSAGE.length).toBeGreaterThan(0);
+});

packages/web/src/features/chat/utils.ts

@@ -176,6 +176,48 @@ export const addLineNumbers = (source: string, lineOffset = 1) => {
     return source.split('\n').map((line, index) => `${index + lineOffset}:${line}`).join('\n');
 }
 
+export const truncateFileContent = (
+    source: string,
+    maxCharacters: number,
+): { content: string; wasTruncated: boolean } => {
+    if (source.length <= maxCharacters) {
+        return { content: source, wasTruncated: false };
+    }
+
+    const cutoff = source.lastIndexOf('\n', maxCharacters);
+    const effectiveCutoff = cutoff > 0 ? cutoff : maxCharacters;
+    const truncated = source.substring(0, effectiveCutoff);
+
+    const totalLines = source.split('\n').length;
+    const includedLines = truncated.split('\n').length;
+
+    return {
+        content: truncated + `\n\n... [truncated: showing ${includedLines} of ${totalLines} lines]`,
+        wasTruncated: true,
+    };
+};
+
+const CONTEXT_WINDOW_ERROR_PATTERNS = [
+    /maximum context length/i,
+    /prompt is too long/i,
+    /context.?length.?exceeded/i,
+    /exceeds? the maximum.*tokens?/i,
+    /token.?limit/i,
+    /request.?too.?large/i,
+    /input.?too.?long/i,
+    /request payload size exceeds/i,
+    /max_tokens/i,
+    /reduce the length/i,
+];
+
+export const isContextWindowError = (errorMessage: string): boolean => {
+    return CONTEXT_WINDOW_ERROR_PATTERNS.some((pattern) => pattern.test(errorMessage));
+};
+
+export const CONTEXT_WINDOW_USER_MESSAGE =
+    'The conversation exceeded the model\'s context window limit. ' +
+    'Try removing some attached files, starting a new conversation, or switching to a model with a larger context window.';
+
 export const createUIMessage = (text: string, mentions: MentionData[], selectedSearchScopes: SearchScope[]): CreateUIMessage<SBChatMessage> => {
     // Converts applicable mentions into sources.
     const sources: Source[] = mentions