A system that investigates, not just searches
Traditional RAG retrieves once and answers. Agentic RAG thinks, retrieves, thinks again, retrieves more, then synthesizes. This allows Claude to follow threads across meetings and emails, identify patterns, and provide analytical insights rather than just relevant quotes.
| Simple RAG | Agentic RAG |
|---|---|
| One retrieval, one answer | Multiple retrievals with reasoning between |
| Finds similar text | Investigates relationships |
| "Here are quotes about X" | "Here's what the data tells us about X" |
| Can't follow threads | Can trace customer journey across sources |
| Static context | Dynamic, expanding context |
User: "Which customers are at risk of churning and why?"
Round 1: Claude searches for complaint-related meetings
→ Finds Jess Main said "when it works, kids love it"
Round 2: Claude gets Jess Main's full history
→ Discovers 3 support tickets in 4 months, last contact 94 days ago
Round 3: Claude cross-references with email threads
→ Finds frustrated tone in recent emails, unresolved issues
Round 4: Claude compares to healthy customer patterns
→ Identifies churn risk factors
Output: Ranked list of at-risk customers with specific evidence and recommendations
| Source | Count | Location |
|---|---|---|
| Meeting transcripts | 747 | raw_meetings/YYYY-MM-DD/*.json |
| Emails | 3,000+ | parsed_emails/emails.jsonl |
| Attachments | 411 unique | parsed_emails/attachments/ |
| Strategic docs | ~3 | reports/MASTER_PLAN.md, reports/cmo_intelligence.md |
Meeting JSON:
{
"id": "123456789",
"title": "Virtual ZTAG Demo with Kris",
"created_at": "2025-06-10T14:30:00Z",
"transcript_data": {
"transcript": [
{
"speaker": {"display_name": "Jess Main"},
"text": "But, like, yeah, when it works, like, the kids love it.",
"start_time": 123.45
}
]
}
}
Email JSONL:
{
"message_id": "<abc123@mail.gmail.com>",
"thread_id": "thread_xyz",
"date": "2025-08-02T10:15:00-07:00",
"from": "jmain@school.edu",
"to": ["sales@ztag.com"],
"subject": "Re: Taggers not connecting",
"body": "Same problem again...",
"is_outbound": false,
"is_reply": true,
"labels": ["Support"]
}
┌─────────────────────────────────────────────────────────────────────────────┐
│ DATA LAYER │
├─────────────────────────────────────────────────────────────────────────────┤
│ │
│ raw_meetings/ parsed_emails/ reports/ │
│ └─ 747 JSONs ├─ emails.jsonl ├─ MASTER_PLAN.md │
│ └─ attachments/ └─ cmo_intelligence.md │
│ │
└──────────────────────────────────┬──────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────────────────┐
│ INDEX LAYER │
├─────────────────────────────────────────────────────────────────────────────┤
│ │
│ ChromaDB (or Pinecone for production) │
│ │
│ Collections: │
│ ├─ meetings (chunks: speaker turns, ~500 tokens each) │
│ │ └─ metadata: meeting_id, date, speaker, segment, customer │
│ │ │
│ ├─ emails (chunks: one email = one chunk) │
│ │ └─ metadata: thread_id, date, from, to, direction, customer │
│ │ │
│ └─ customers (derived: profiles built from both sources) │
│ └─ metadata: name, segment, first_contact, last_contact, health_score │
│ │
│ Embedding model: text-embedding-3-small or Voyage AI │
│ │
└──────────────────────────────────┬──────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────────────────┐
│ TOOL LAYER │
├─────────────────────────────────────────────────────────────────────────────┤
│ │
│ Tools available to Claude: │
│ │
│ search_meetings(query, segment?, limit?) │
│ └─ Semantic search over meeting transcripts │
│ │
│ search_emails(query, direction?, limit?) │
│ └─ Semantic search over email threads │
│ │
│ get_customer_history(customer_name, include_transcripts?) │
│ └─ ALL meetings + emails for a customer, chronological │
│ │
│ get_full_meeting(meeting_id) │
│ └─ Complete transcript for deep reading │
│ │
│ get_email_thread(thread_id) │
│ └─ All emails in thread, chronological │
│ │
│ list_customers(segment?, no_contact_days?) │
│ └─ Filter customers by segment or recency │
│ │
└──────────────────────────────────┬──────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────────────────┐
│ ORCHESTRATOR (Claude) │
├─────────────────────────────────────────────────────────────────────────────┤
│ │
│ System prompt includes: │
│ - ZTAG business context (from MASTER_PLAN.md) │
│ - Segment definitions (institutional vs professional) │
│ - Investigation patterns (how to analyze churn, satisfaction, etc.) │
│ │
│ Agentic loop: │
│ 1. Receive user question │
│ 2. Think: what do I need to investigate? │
│ 3. Call tools (can be multiple per round) │
│ 4. Receive results, think again │
│ 5. Repeat until confident (max 5 rounds) │
│ 6. Synthesize final answer with evidence │
│ │
└─────────────────────────────────────────────────────────────────────────────┘
Script: scripts/build_index.py
# Estimated indexing cost
# Meetings: 747 × ~20 turns × ~200 tokens = ~3M tokens = $0.06
# Emails: 3,000 × ~500 tokens = ~1.5M tokens = $0.03
# Total one-time cost: ~$0.10
Script: scripts/build_customer_index.py
Link customers across sources:
Script: scripts/rag_tools.py
Implement the 6 tools with proper ChromaDB queries:
def search_meetings(query: str, segment: str = None, limit: int = 10) -> list:
"""Semantic search over meeting transcripts."""
where = {"segment": segment} if segment else None
results = meetings_collection.query(
query_texts=[query],
n_results=limit,
where=where
)
return format_results(results)
def get_customer_history(customer_name: str) -> dict:
"""Get ALL meetings and emails for a customer."""
# Fuzzy match customer name
customer = resolve_customer(customer_name)
# Get all interactions
meetings = get_meetings_by_customer(customer.id)
emails = get_emails_by_customer(customer.id)
return {
"customer": customer.name,
"segment": customer.segment,
"first_contact": customer.first_contact,
"last_contact": customer.last_contact,
"meetings": meetings,
"emails": emails,
"support_tickets": len([e for e in emails if "support" in e.labels])
}
Script: scripts/agentic_rag.py
from anthropic import Anthropic
TOOLS = [
{"name": "search_meetings", ...},
{"name": "search_emails", ...},
{"name": "get_customer_history", ...},
{"name": "get_full_meeting", ...},
{"name": "get_email_thread", ...},
{"name": "list_customers", ...},
]
SYSTEM_PROMPT = """
You are ZTAG's business intelligence analyst. You have access to:
- 747 customer meeting transcripts (demos, training, support calls)
- 3,000+ sales emails
- Customer profiles across two segments: Institutional (schools) and Professional (GameTruck)
Your job is to INVESTIGATE questions by using the available tools. Don't guess—retrieve the data.
For complex questions:
1. Start broad (search for relevant mentions)
2. Go deep (get full history for interesting customers)
3. Cross-reference (meetings + emails for same customer)
4. Synthesize (identify patterns, not just examples)
"""
def agentic_query(question: str, max_rounds: int = 5) -> str:
messages = [{"role": "user", "content": question}]
for round_num in range(max_rounds):
response = client.messages.create(
model="claude-sonnet-4-20250514",
max_tokens=4096,
system=SYSTEM_PROMPT,
tools=TOOLS,
messages=messages
)
if response.stop_reason == "tool_use":
# Execute tools and continue
tool_results = execute_all_tools(response.content)
messages.append({"role": "assistant", "content": response.content})
messages.append({"role": "user", "content": tool_results})
else:
# Done - return final answer
return response.content[0].text
return "Max rounds reached."
Option A: CLI
python scripts/ask.py "Which customers are at risk of churning?"
Option B: API (FastAPI)
@app.post("/ask")
async def ask(question: str):
answer = agentic_query(question)
return {"answer": answer}
Option C: Slack Bot
| Component | One-time | Per Query |
|---|---|---|
| Embedding (indexing) | ~$0.10 | - |
| ChromaDB storage | Free (local) | - |
| Claude Sonnet queries | - | ~$0.02-0.05 |
| Estimated monthly (100 queries/day) | - | ~$60-150 |
anthropic>=0.18.0
chromadb>=0.4.0
openai>=1.0.0 # for embeddings
python-dotenv
meetings/
├── scripts/
│ ├── build_index.py # Phase 1: Index meetings + emails
│ ├── build_customer_index.py # Phase 2: Entity resolution
│ ├── rag_tools.py # Phase 3: Tool implementations
│ ├── agentic_rag.py # Phase 4: Orchestrator
│ └── ask.py # Phase 5: CLI interface
├── chroma_db/ # Vector database storage
│ ├── meetings/
│ ├── emails/
│ └── customers/
└── docs/
└── AGENTIC_RAG_ARCHITECTURE.md # This file
Document version: 1.0
Created: February 2026
Author: Claude + Quan