Optimize AI Accuracy: Fix Data Issues to Reduce Hallucinations

Optimize AI Accuracy: Fix Data Issues to Reduce HallucinationsOptimize AI Accuracy: Fix Data Issues to Reduce Hallucinations

The network for creativity

Join 1.25M professional creatives like you

Connect with clients, get discovered, and run your business 100% commission-free

Creatives on Contra have earned over $150M and we are just getting started

Back to feedPost

Mark Bright

• 1d

RAG is not the bottleneck. Your data is.

Everyone's tweaking embeddings, swapping models, A/B testing prompt formats. Then they ship. Then the bot still hallucinates.

Here's where I usually see things break:

🚫 What people debug • "Maybe gpt-4o is the wrong model" • "Try a different embedding model" • "Increase top_k to 20" • "Add a reranker, that'll fix it"

✅ What actually fixes it • Chunk by meaning, not by 500 tokens • Keep the source structure (headings, tables, lists) • Write summaries of long sections, embed those too • Drop duplicates, drop legal footers, drop nav menus • Test on real user questions, not your own

Most "AI accuracy" problems are data problems pretending to be model problems.

Build the pipeline properly once, and the same model gives you 2x better answers.

If you're stuck on this, drop a comment with your case. Happy to debug for free, I learn from each one.

aiengineering rag chatbot

Lifa Yende

• May 24

This is the difference between a generic WhatsApp bot and what I build. Left side: a standard menu bot that breaks the moment a patient asks a real question. Right side: NowNow AI's conversational WhatsApp agent — it answers pricing instantly, handles objections, checks medical aid coverage, follows up the next day unprompted, and books the appointment with full confirmation details. No menus, no dead ends, no "Sorry, I didn't understand that." Built on Cloudflare Workers with Claude-powered conversation memory, Meta WhatsApp Business API, and Make.com orchestration. Every message you see here was handled autonomously — zero human intervention from first enquiry to confirmed booking while still keeping a conversational and human tone

Conversational AI Development WhatsApp AI Chatbot Development Business Workflow Automation Cloudflare Workers N8N

Sabrina Polanco

max

• May 21

Which AI model would actually survive a real ad agency brief?

I put Kling 3.0 and Seedance 2.0 head-to-head with the exact same frames. Same prompt. Same creative direction.

And honestly, choosing between models is starting to feel a lot closer to choosing a director than choosing a tool.

Because each one interprets motion, realism, pacing, and atmosphere completely differently.

Prompt’s in the comments if you want to test it with your own product.

They both still need some extra editing and work, but I definitely have a favorite.

Who’s winning this one for you?

aivideo kling3 seedance2

29 voted

39%

45 voted

61%

74 votes

Closed

Sabrina Polanco

max

• May 21

The starting frame vs ending frame.

I generated both frames with Nano Banana 2 and used them as the transition points for the video generation.

I wanted it to feel like one continuous world: floating tension → gravity switches on → full liquid impact

Luis Rendon

• May 24

Production WhatsApp AI agent built for a real client (PixelPages). The bot handles customer inquiries 24/7 using a RAG pipeline — it searches a vector knowledge base, generates context-aware responses via OpenAI, and automatically escalates to a human agent when it hits the boundaries of its knowledge. Self-hosted on Oracle Cloud with Docker, orchestrated with n8n, and running in production across multiple tenants with isolated data per client.