Agentic Outreach Pipeline — Demo Guide

Quick-start guide to run and test every feature of the pipeline.

Prerequisites

# Python 3.13 required (ChromaDB does not support 3.14)
python3 --version   # should show 3.13.x

# Anthropic API key set in .env
cat .env   # should contain ANTHROPIC_API_KEY=sk-ant-...

Setup (one-time)

cd agentic-outreach-pipeline
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Test 1: Verify Imports and Configuration

python -c "
import sys; sys.path.insert(0, 'src')
from orchestrator import Orchestrator
from knowledge.product_loader import COMPANY_CONFIG, PRODUCTS_CONFIG
from context import ContextManager

print(f'Company: {COMPANY_CONFIG[\"name\"]}')
print(f'Products: {list(PRODUCTS_CONFIG[\"products\"].keys())}')
print(f'Case studies: {len(PRODUCTS_CONFIG[\"case_studies\"])}')
print('All imports OK')
"

Expected: Company name, product keys, case study count, and "All imports OK".

Test 2: View Execution Plan (No API Calls)

python src/main.py plan "Mueller Automotive GmbH, Germany"

Expected: Structured log output showing the pipeline plan:

pipeline_plan                  target='Mueller Automotive GmbH'
plan_step                      agent=researcher  criticality=required   group=0
plan_step_parallel             agents='Analyst + Architect'             group=1
plan_step                      agent=analyst     criticality=optional   group=1
plan_step                      agent=architect   criticality=required   group=1
plan_step                      agent=scorer      criticality=optional   group=2
plan_step                      agent=writer      criticality=required   group=3

This verifies the orchestrator and models work without spending any API credits.

Test 3: Verify Knowledge Base Query (ChromaDB Semantic Search)

python -c "
import sys; sys.path.insert(0, 'src')
from tools.knowledge_query import KnowledgeQueryTool

tool = KnowledgeQueryTool()
result = tool.run(query='energy monitoring for stamping factories')
print(result[:500])
"

Expected: Relevant product knowledge matching the query — product features, case studies, or ideal customer data. Results include relevance scores from ChromaDB vector similarity search.

First run note: ChromaDB downloads the all-MiniLM-L6-v2 embedding model (~80MB) on first use. Subsequent runs are instant.

Test 3b: Verify Semantic Search (RAG Verification)

python -c "
import sys; sys.path.insert(0, 'src')
from tools.knowledge_query import KnowledgeQueryTool

tool = KnowledgeQueryTool()

# Semantic search: no shared words with 'energy monitoring' but should still match
result = tool.run(query='power consumption analysis for metal forming')
print('=== Semantic match test ===')
print(result[:500])
print()

# Verify ChromaDB collections
from knowledge.store import get_knowledge_store
store = get_knowledge_store()
print(f'Store type: {type(store).__name__}')
print(f'ChromaDB initialized: {hasattr(store, \"_client\")}')
"

Expected: Semantic search returns relevant product knowledge about energy/power monitoring even though the query uses different words ("power consumption analysis" vs "energy monitoring"). Store type should be VectorKnowledgeStore.

Test 4: Verify Web Scraper

python -c "
import sys; sys.path.insert(0, 'src')
from tools.web_scraper import WebScraperTool

scraper = WebScraperTool()
result = scraper.run(url='https://httpbin.org/html')
print(result[:300])
"

Expected: Clean text from the page. Errors return strings like Error: Could not connect..., not crashes.

Test 5: Generate a Proposal (Main Test)

python src/main.py proposal "Koelle GmbH, Germany"

What to watch for:

▸ Researcher — multi-turn agentic research starts
Tool calls: search_web(...), query_knowledge_base(...), scrape_company_website(...)
Turn progress: Turn 1/5, Turn 2/5, etc.
✓ Done — researcher complete with token counts
▸ Analyst + ▸ Architect — running in parallel
▸ Scorer — deal estimation
▸ Writer — proposal generation
Pipeline summary table with total tokens and duration

Output: Check outputs/proposal_Koelle_GmbH_*.md for the full proposal.

Test 6: Try Example Prospects

python src/main.py --example

Pick 1, 2, or 3 from the menu. Each runs the full proposal pipeline with a pre-configured company.

#	Prospect	Industry
1	Mueller Automotive GmbH, Germany	Auto parts stamping, 150 presses
2	Pacific Brass & Copper, USA	Copper fittings, 40+ machines
3	Vina Precision Parts, Vietnam	Electronics stamping, 80 presses

Test 7: Interactive Mode

python src/main.py --interactive

Type a company description (press Enter twice to submit):

Samsung SDI, South Korea
Battery cell manufacturer for EV market
Large-scale production with 200+ machines
ESG compliance critical for European automotive customers

Test 8: Search & Outreach Pipeline

python src/main.py search "metal stamping companies Germany"

What happens:

DuckDuckGo finds companies
Contact info extracted (emails, phones)
Deal sizes estimated (structured JSON via tool_choice)
Rich table displayed — pick companies to pursue
Full pipeline runs for each selected prospect (researcher + architect + writer)
Cold emails written and previewed
Confirm to send via Gmail (or skip)
Results saved to outputs/outreach_*.json

Try also:

python src/main.py search "automotive parts manufacturer Japan"
python src/main.py search "copper fittings factory USA"
python src/main.py search "injection molding company Vietnam"

Test 9: Streamlit Web UI

streamlit run app.py

Open http://localhost:8501. The UI has two tabs:

Tab 1: Single Proposal

Select an example prospect from the dropdown (or enter custom input)
Click "Run Sales Agent Pipeline" — watch all agents run with streaming progress
Expand Research Brief and Solution Mapping to review intermediate outputs
Download the proposal as Markdown
Click "Generate Cold Email" — runs the Deal Estimator + Email Writer
Edit the email, enter a recipient, and click "Send Email" (requires Gmail in .env)

Tab 2: Prospect Search

Enter a search query (e.g. "metal stamping companies Germany")
Set max results and click "Search"
Review the prospect table (company, industry, email, estimated deal, category)
Select companies and click "Generate Detailed Proposals"
Full pipeline runs per company: Researcher → Architect → Writer
Review proposals, emails, and research for each company
Download or send emails directly from the UI

Test 10: Run Tests

pytest tests/ -v

Expected: All tests pass. Tests use mocked API calls and don't require an Anthropic API key.

Output Files

After running tests, check:

ls -la outputs/

File Pattern	From
`proposal_*.md`	Proposal mode — full Markdown proposal
`pipeline_*.json`	Proposal mode — debug output with agent results
`trace_*.json`	Proposal mode — per-agent timing and cost
`outreach_*.json`	Search mode — prospects, deals, emails

Troubleshooting

Issue	Fix
`ModuleNotFoundError`	`source venv/bin/activate && pip install -r requirements.txt`
`credit balance is too low`	Add credits at https://console.anthropic.com/settings/billing
`No companies found` (search)	Try more specific terms: "CNC machining factory Vietnam"
Rate limited during pipeline	Built-in exponential backoff handles this (3 retries, 10-40s waits)
Researcher falls back to legacy	API error during multi-turn research — check terminal logs
Gmail authentication failed	Use a Google App Password, not your regular password
ChromaDB import error	ChromaDB requires Python <=3.13. Recreate venv: `python3.13 -m venv venv`
Slow first knowledge query	ChromaDB downloads all-MiniLM-L6-v2 (~80MB) on first run. Cached at `~/.cache/chroma/`
Stale knowledge results	Delete `chroma_data/` directory and restart — it auto-recreates and re-seeds from YAML config
`chromadb.errors` on startup	Delete `chroma_data/` and let it regenerate: `rm -rf chroma_data/`

Quick Smoke Test (30 seconds)

Verify everything works without spending API credits:

python -c "
import sys; sys.path.insert(0, 'src')

# 1. Models
from models import PipelineResult, DealEstimate, ContextPacket
print('Models OK')

# 2. Knowledge
from knowledge.product_loader import COMPANY_CONFIG, get_full_product_context
ctx = get_full_product_context()
print(f'Knowledge OK ({len(ctx)} chars, {len(ctx.split(chr(10)))} lines)')

# 3. Tools
from tools.web_search import WebSearchTool
from tools.web_scraper import WebScraperTool
from tools.knowledge_query import KnowledgeQueryTool
print(f'Tools OK (3 tools loaded)')

# 4. ChromaDB vector store
from knowledge.store import get_knowledge_store
store = get_knowledge_store()
print(f'ChromaDB OK (store type: {type(store).__name__})')

# 5. Knowledge query (semantic search)
kb = KnowledgeQueryTool()
result = kb.run(query='defect detection')
print(f'KB query OK ({len(result)} chars)')

# 6. Orchestrator
from orchestrator import Orchestrator
o = Orchestrator(interactive=True)
plan = o.plan('Test Company, Germany')
print(f'Orchestrator OK ({len(plan.steps)} steps)')

# 7. Context
from context import ContextManager
cm = ContextManager()
tokens = cm.count_tokens('hello world')
print(f'Context OK ({tokens} tokens)')

print()
print('All systems operational')
"

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Agentic Outreach Pipeline — Demo Guide

Prerequisites

Setup (one-time)

Test 1: Verify Imports and Configuration

Test 2: View Execution Plan (No API Calls)

Test 3: Verify Knowledge Base Query (ChromaDB Semantic Search)

Test 3b: Verify Semantic Search (RAG Verification)

Test 4: Verify Web Scraper

Test 5: Generate a Proposal (Main Test)

Test 6: Try Example Prospects

Test 7: Interactive Mode

Test 8: Search & Outreach Pipeline

Test 9: Streamlit Web UI

Tab 1: Single Proposal

Tab 2: Prospect Search

Test 10: Run Tests

Output Files

Troubleshooting

Quick Smoke Test (30 seconds)

FilesExpand file tree

DEMO.md

Latest commit

History

DEMO.md

File metadata and controls

Agentic Outreach Pipeline — Demo Guide

Prerequisites

Setup (one-time)

Test 1: Verify Imports and Configuration

Test 2: View Execution Plan (No API Calls)

Test 3: Verify Knowledge Base Query (ChromaDB Semantic Search)

Test 3b: Verify Semantic Search (RAG Verification)

Test 4: Verify Web Scraper

Test 5: Generate a Proposal (Main Test)

Test 6: Try Example Prospects

Test 7: Interactive Mode

Test 8: Search & Outreach Pipeline

Test 9: Streamlit Web UI

Tab 1: Single Proposal

Tab 2: Prospect Search

Test 10: Run Tests

Output Files

Troubleshooting

Quick Smoke Test (30 seconds)