Scenario-Aware JSON Re-Contextualization

A LangGraph-based agentic workflow that transforms JSON documents to match a new scenario while preserving structure and locked fields.

What It Does

Takes an input JSON with multiple scenarios and re-contextualizes ALL content to match a newly selected scenario - including company names, brands, person names, emails, URLs, KPIs, pricing, and narrative text.

Guarantees:

• Output JSON has identical structure to input (same keys, nesting, array lengths)
• scenarioOptions field remains byte-for-byte identical (the only locked field)
• Everything else is rewritten for the new scenario
• Brands updated: HarvestBowls → TrendWave
• Person names updated: Mark Caldwell → Sarah Chen
• Emails updated: mark@harvestbowls.com → sarah@trendwave.com
• Industry-adaptive pricing: $2-$3 (food) → $25-$35 (fashion)
• Zero residual references from old scenario

Performance

Metric	Value
Total Runtime	7-8 seconds
LLM Calls	~42 calls
Texts Processed	172
Retries	0-1 typically

Key Features

Feature	Description
LLM Entity Extraction	Parallel extraction of brands, people, emails from both scenarios
Few-Shot Learning	Dynamic examples generated from scenario mapping
Industry-Adaptive Pricing	Automatic scaling of prices to fit target industry
Token-Aware Batching	Batches sized by token count (~600 tokens/batch)
HTML Splitting	Large HTML split at <h2>, <h3> boundaries for parallel processing
Comprehensive Validation	Checks brands, person names, email domains, industry consistency

Usage

# Install dependencies
pip install -r requirements.txt
# Set OpenAI API keyexport OPENAI_API_KEY="your-key-here"# Run re-contextualization (by scenario index)
python -m src.main --input POC_sim_D.json --scenario 3
# Run with scenario string
python -m src.main --input POC_sim_D.json --scenario "A mid-market fashion retailer..."# List available scenarios
python -m src.main --input POC_sim_D.json --list-scenarios

Output

output/
├── output.json # Re-contextualized JSON
└── validation_report.json # Validation results + runtime stats

Validation Report Structure

{
"schemaMatch": true,
"lockedFieldsUntouched": true,
"changedFieldPaths": [
"topicWizardData.lessonInformation.lesson",
"topicWizardData.workplaceScenario.background.organizationName",
"topicWizardData.simulationFlow[0].children[1].data.email.body",
"... (91 paths total)"
],
"scenarioConsistency": "OK",
"failedPaths": [],
"runtimeStats": {
"latency_ms": 7616,
"numRetries": 0,
"numLLMCalls": 42
}
}

---

Architecture

High-Level Flow

┌──────────────────────────────────────────────────────────────────────┐
│ 1. PARSE & EXTRACT │
│ • Pull ALL strings from JSON (172 text values) │
│ • Only skip: scenarioOptions (locked field) │
│ • LLM Entity Extraction (parallel): │
│ - OLD: company=HarvestBowls, competitor=Nature's Crust │
│ - NEW: company=TrendWave, competitor=ChicStyles │
│ - Person names, email domains, promotions │
└──────────────────────────────────────────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────────────────────┐
│ 2. GENERATE (LLM Rewrite) │
│ • Few-shot examples from entity mapping │
│ • Industry-adaptive pricing rules │
│ • 40 parallel batches × ~600 tokens = 5-6s │
│ • Large HTML split into sections, reassembled after │
└──────────────────────────────────────────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────────────────────┐
│ 3. ASSEMBLE │
│ • Deep copy original JSON structure │
│ • Inject rewritten texts at original paths │
│ • Reassemble split HTML sections │
│ • Restore locked scenarioOptions │
└──────────────────────────────────────────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────────────────────┐
│ 4. VALIDATE (LLM-Extracted Entities) │
│ • Schema match? (same structure) │
│ • Locked fields intact? (scenarioOptions unchanged) │
│ • No old entities? (brands, people, emails, promotions) │
│ • If leakage found → targeted repair (max 3 retries) │
└──────────────────────────────────────────────────────────────────────┘

LangGraph Workflow

┌─────────────────────┐ ┌─────────────────────┐ ┌─────────────────┐
│ Parse & Extract │────▶│ Generate │────▶│ Assemble │
│ + LLM Entity Ext. │ │ (parallel batches) │ │ │
│ (parallel) │ │ + Few-Shot Learn │ │ │
└─────────────────────┘ └─────────────────────┘ └─────────────────┘
│
┌─────────────────┐ │
│ Repair │◀────────────────┤ (if validation fails)
│ (retry failed) │ │
└─────────────────┘ ▼
│ ┌─────────────────┐
└──────────────────▶│ Validate │────▶ END
│ (LLM entities) │
└─────────────────┘

5-Node Pipeline

Node	What It Does
Parse	Extract ALL strings, LLM entity extraction (parallel for old/new)
Generate	Rewrite with few-shot examples + industry-adaptive pricing
Assemble	Inject texts back, reassemble HTML sections, restore locked fields
Validate	Check schema, locked fields, entity leakage (brands, people, emails)
Repair	Retry only failed texts with specific feedback (max 3 retries)

Project Structure

context/
├── src/
│ ├── __init__.py # Package marker
│ ├── config.py # Model name, API settings, batching constants
│ ├── workflow.py # LangGraph nodes + graph definition (~1000 lines)
│ └── main.py # CLI entry point
├── output/ # Generated outputs
├── requirements.txt
└── POC_sim_D.json # Sample input

---

Key Design Decisions

1. LLM-Based Entity Extraction

Instead of brittle regex patterns, we use LLM to semantically extract entities from scenarios:

# LLM extracts these entities (run in parallel for old & new scenarios):
{
"primary_company": "HarvestBowls", # → "TrendWave""competitor": "Nature's Crust", # → "ChicStyles""promotion_type": "$1 value menu", # → "Buy One Get One Free""person_names": ["Mark Caldwell", "Emily Carter"],
"email_domains": ["harvestbowls.com"]
}

Benefits:

• Handles apostrophes: Nature's Crust ✓
• Finds names not in scenario text (from JSON context)
• Infers email domains from company names
• No regex maintenance required

2. Few-Shot Learning

Dynamic examples generated from entity mapping teach the LLM the transformation pattern:

TRANSFORMATION EXAMPLES:
─────────────────────────
Brand: "HarvestBowls" → "TrendWave"
Competitor: "Nature's Crust" → "ChicStyles"
Person: "Mark Caldwell" → [Generate new name like "Sarah Chen"]
Email: "mark@harvestbowls.com" → "sarah@trendwave.com"
Promotion: "$1 value menu" → "Buy One Get One Free promotion"
Industry: "fast food" → "fashion retail"

3. Industry-Adaptive Pricing

The LLM prompt includes rules to scale prices appropriately:

PRICING RULES:
• Food → Fashion: $2-$3 items → $25-$35 items
• Fashion → Food: $50-$70 items → $5-$7 items
• Scale ALL dollar amounts proportionally
• Keep percentages and relative comparisons the same

Example Transformation:

BEFORE (Food): "Launch a HarvestBites menu with items at $2-$3"
AFTER (Fashion): "Launch a TrendWave collection with items at $25-$35"

4. Comprehensive Validation

Validation now checks for ALL old scenario markers using LLM-extracted entities:

Check	What It Verifies
Schema Match	Same keys, nesting, array lengths
Locked Fields	scenarioOptions byte-for-byte identical
Brand Leakage	No old company/competitor names
Person Name Leakage	No old person names (LLM-extracted)
Email Domain Leakage	No old email domains
Promotion Leakage	No old promotion references

5. Only Lock scenarioOptions

Per requirements, only ONE field is locked:

scenarioOptions → MUST be byte-for-byte identical
Everything else → CAN change to match new scenario

6. Token-Aware Parallel Batching

Parameter	Value
Batch Size	Dynamic ~600 tokens per batch
Large Text (HTML)	Split at <h2>, <h3> boundaries
Parallelism	All batches run concurrently
Total Batches	~40 batches
Max Batch Time	~5 seconds

7. Targeted Repair

If validation finds leakage:

1. Identify which specific texts failed
2. Reset only those texts
3. Re-generate with specific feedback
4. Loop back to validation

Max 3 retries, then output with warning.

---

Transformation Example

Input Scenario (Food Industry):

A strategy team at HarvestBowls is facing a drop in foot traffic after
Nature's Crust introduced a $1 value menu. As a business consultant,
learners must analyze the market shake-up...

Output Scenario (Fashion Industry):

A mid-market fashion retailer, TrendWave, is losing customers to
ChicStyles' aggressive Buy One Get One Free promotion. As a business
consultant, learners must analyze the market shake-up...

What Changes:

Element	Before	After
Company	HarvestBowls	TrendWave
Competitor	Nature's Crust	ChicStyles
Promotion	$1 value menu	Buy One Get One Free
Manager	Mark Caldwell	Sarah Chen
Email	mark@harvestbowls.com	sarah@trendwave.com
Pricing	$2-$3 items	$25-$35 items
Industry Terms	fast food, menu, organic	fashion retail, apparel

---

Requirements

• Python 3.10+
• OpenAI API key
• Dependencies: langgraph, langchain, langchain-openai, pydantic, python-dotenv

Cost Estimate

Per re-contextualization run (~172 text values):

Component	Calculation	Cost
Entity extraction	~2K × $0.10/1M	$0.0002
Input tokens (rewrite)	~18K × $0.10/1M	$0.0018
Output tokens	~12K × $0.40/1M	$0.0048
Total		~$0.007/run

At scale: ~$7 per 1,000 documents processed.

---

Verification Checklist

Run this to verify all requirements are met:

python -c "import jsonwith open('output/output.json') as f: output = json.load(f)with open('output/validation_report.json') as f: report = json.load(f)print('1. Schema Match:', '✅' if report['schemaMatch'] else '❌')print('2. Locked Fields:', '✅' if report['lockedFieldsUntouched'] else '❌')print('3. Consistency:', '✅' if report['scenarioConsistency'] == 'OK' else '❌')print('4. Runtime:', f\"{report['runtimeStats']['latency_ms']}ms\")print('5. Changed Fields:', len(report['changedFieldPaths']))"

Expected output:

1. Schema Match: ✅
2. Locked Fields: ✅
3. Consistency: ✅
4. Runtime: ~7600ms
5. Changed Fields: 91