github/node-DeepResearch
1
0
Fork 0
mirror of https://github.com/jina-ai/node-DeepResearch.git synced 2026-03-22 07:29:35 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix: update token counting to use content length estimation

Browse Source 
Co-Authored-By: Han Xiao <han.xiao@jina.ai>
This commit is contained in:
Devin AI
2025-02-11 09:28:34 +00:00
parent c4639a2e92
commit 7e698639b1
1 changed files with 14 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
src/server.ts
View File

@@ -193,10 +193,13 @@ app.post('/v1/chat/completions', (async (req: Request, res: Response) => {
actionTracker: new ActionTracker()
};
// Track prompt tokens for the initial message
// Use Vercel's token counting convention - 1 token per message
const messageTokens = body.messages.length;
context.tokenTracker.trackUsage('agent', messageTokens, TOKEN_CATEGORIES.PROMPT);
// Track prompt tokens for each message using actual content length
for (const message of body.messages) {
// Estimate tokens using character count / 4 as a rough approximation
// This will be replaced with actual Gemini tokenizer in a future update
const estimatedTokens = Math.ceil(Buffer.byteLength(message.content, 'utf-8') / 4);
context.tokenTracker.trackUsage('agent', estimatedTokens, TOKEN_CATEGORIES.PROMPT);
}
// Add this inside the chat completions endpoint, before setting up the action listener
const streamingState: StreamingState = {
@@ -316,13 +319,14 @@ app.post('/v1/chat/completions', (async (req: Request, res: Response) => {
// Track tokens based on action type
if (result.action === 'answer') {
// Track accepted prediction tokens for the final answer using Vercel's convention
const answerTokens = 1; // Default to 1 token per answer
context.tokenTracker.trackUsage('evaluator', answerTokens, TOKEN_CATEGORIES.ACCEPTED);
// Track tokens for the final answer using content length estimation
const content = result.action === 'answer' ? buildMdFromAnswer(result) : result.think;
const estimatedTokens = Math.ceil(Buffer.byteLength(content, 'utf-8') / 4);
context.tokenTracker.trackUsage('evaluator', estimatedTokens, TOKEN_CATEGORIES.ACCEPTED);
} else {
// Track rejected prediction tokens for non-answer responses
const rejectedTokens = 1; // Default to 1 token per rejected response
context.tokenTracker.trackUsage('evaluator', rejectedTokens, TOKEN_CATEGORIES.REJECTED);
// Track tokens for non-answer responses using content length estimation
const estimatedTokens = Math.ceil(Buffer.byteLength(result.think, 'utf-8') / 4);
context.tokenTracker.trackUsage('evaluator', estimatedTokens, TOKEN_CATEGORIES.REJECTED);
}
if (body.stream) {