Full breakdown with implementation details: https://www.planetnocode.com/bubble-tutorials/3-ways-save-ai-chat-history-bubble-io-apps

Bonus: Hybrid search (semantic + keyword) >> semantic alone.
Semantic search fails on proper nouns like company names. But hybrid search is only available with Pinecone's native embeddings—another reason to use them.

1. Call OpenAI API for embeddings first
2. Send vectors to Pinecone
3. Repeat for every query

If I'd used Pinecone's native embeddings, I could just send TEXT directly. Save yourself this extra step.

Key lesson I learned the hard way with Pinecone:

I used OpenAI's embeddings models (text-3-small/large), which means I have to:

METHOD 3: RAG with Pinecone
Vector databases enable semantic search—understanding "cat" and "kitten" are related, not just exact spelling matches.
You convert each message pair into embeddings (numbers), store them in Pinecone, then retrieve only relevant history based on meaning

METHOD 2: Follow-up Prompt (my favorite for cost)
Use an expensive model for responses, then run a cheap background model ($0.0002) to distill the conversation into a structured profile:

- Role
- Task state
- Learner progress
- Decision points
- Next steps

METHOD 1: Context Window
Yes, we now have 200K-1M token limits. But three problems:

1. Recent research shows AI loses track of info in the MIDDLE of conversations
 2. Costs compound exponentially (12 messages, then 14, then 16...)

3. Claude API has hard message limits (~1000)

The future of support might not be human versus AI. It might be human plus AI.
#AICoaching #FutureOfWork #HumanConnection

For millions without access to human coaching, this could mean 24/7 affordable support.
AI may not be ready to replace deep therapeutic relationships yet.

But the question isn't whether it will happen. It's how quickly we adapt to working alongside these tools rather than competing with them.

This points to a fascinating division of labour emerging.

AI handling structured, goal-focused support that follows established frameworks.
Humans managing the complex, adaptive work requiring cultural sensitivity and emotional nuance.

The research revealed something unexpected about our relationship with AI. We don't need to build rapport with machines like we do with humans. What matters most is whether we believe the technology actually works.

Students felt psychologically safe with AI. They shared personal information without fear of judgement.

But here's where it gets interesting.

AI only worked for narrow targets. It couldn't improve broader measures like resilience or overall wellbeing, which human coaches influenced significantly.