Customer Support POC

AI-powered customer support demo using the Neocortex Memory SDK (tinyhumansai). Built to demonstrate how memory-backed AI agents handle real support scenarios for a fictional SaaS product, CloudSync Pro.

Use Cases

#	Page	What it shows
1	Seed Data	Load/clear sample data into Neocortex namespaces
2	Community Knowledge	Semantic search over forum/Discord solutions — no curation needed
3	Ticket Resolution	Find similar past tickets + reinforce good solutions (re-ingest with fresh timestamp)
4	KB Evolution	Deprecate outdated articles, reinforce current ones, add new content
5	Customer Context	Personalized chat — agent recalls past interactions, plan details, preferences

Architecture

┌─────────────────────────────────────────────────┐
│                  Streamlit UI                    │
│  (Seed Data | Community | Tickets | KB | Chat)  │
└─────────────────┬───────────────────────────────┘
                  │
        ┌─────────┴─────────┐
        │  support_agent.py │  ← multi-namespace recall + LLM
        └─────────┬─────────┘
                  │
    ┌─────────────┴─────────────┐
    │  Neocortex Memory SDK     │
    │  (tinyhumansai)           │
    │                           │
    │  Namespaces:              │
    │  ├─ community-solutions   │
    │  ├─ resolved-tickets      │
    │  ├─ knowledge-base        │
    │  └─ customer:{id}         │
    └─────────────┬─────────────┘
                  │
          ┌───────┴───────┐
          │  OpenAI API   │  ← response generation via recall_with_llm
          └───────────────┘

Quick Start

# 1. Clone
git clone https://github.com/tinyhumansai/example-customer-support.git
cd example-customer-support

# 2. Install dependencies
pip install tinyhumansai streamlit python-dotenv

# 3. Configure API keys
cp .env.example .env

Open .env and fill in your keys:

Variable	Where to get it
TINYHUMANS_TOKEN	app.tinyhumans.ai → Settings → API Keys → Generate
TINYHUMANS_MODEL_ID	Same page — copy the Model ID for your project
OPENAI_API_KEY	platform.openai.com/api-keys

Note: OPENAI_API_KEY is only needed for pages that generate LLM responses (pages 2–5). Seeding data (page 1) works without it.

# 4. Run
streamlit run app.py

Then:

1. Go to Seed Data page and click Seed All Data
2. Explore each use-case page

Key Concepts

Reinforcement = Re-ingestion

The SDK upserts on (namespace, key). Re-ingesting an item with a fresh updated_at timestamp naturally reinforces it — the item becomes more prominent in recall results.

Decay = Deletion

delete_memory removes outdated content. In production, This could automate this based on article age, negative feedback count, or content freshness scores.

Namespace Isolation

Each use case has its own namespace. Customer data is further isolated per customer (customer:acme-corp). The support agent can recall across all namespaces simultaneously for comprehensive answers.

Sample Data

All in data/:

• community_posts.json — 15 forum/Discord solutions (sync, API, SSO, mobile, etc.)
• resolved_tickets.json — 10 historical tickets with full issue→diagnosis→resolution
• kb_articles.json — 8 KB articles including deprecated v1 and current v2 versions
• customers.json — 3 customers (Acme Corp, TechStart Inc, Global Media) with interaction histories

Next Steps

• LangChain/LangGraph integration — the SDK supports TinyHumanStore for LangChain vector store compatibility
• Automated decay — scheduled jobs to deprecate low-performing content
• Feedback loop — track resolution success rates to auto-reinforce effective solutions
• Multi-language support — the SDK handles embedding-based recall, so multilingual content works out of the box