1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621# LLM Content Generation Evaluation App - Full Specification
## Project Background & Goals
### Overview
We're building a web application to systematically evaluate which LLM (OpenAI, Anthropic, Google) and which prompting strategy produces the best marketing content. The user is a marketer who wants to test different models and prompting approaches to find what generates content most aligned with their personal writing style.
### Core Objectives
1. Test 3 top-tier LLM providers (OpenAI GPT-4, Anthropic Claude, Google Gemini)
2. Compare 2 prompting strategies (Structured vs Example-based)
3. Evaluate across 3 content types (Blog intros, LinkedIn posts, Company announcements)
4. Support single-person evaluation with blind testing to reduce bias
5. Analyze results to identify best model/prompt combinations for each content type
### Experimental Design
- **Total generations**: 36 (3 models × 2 prompt strategies × 6 tasks)
- **Evaluation method**: Human judgment on style, coherence, and publishability
- **Key innovation**: Using the evaluator's own writing samples as style references
## Technical Specification
### Tech Stack
- **Backend**: FastAPI
- **Database**: SQLite with SQLAlchemy ORM
- **Frontend**: HTML/JavaScript with Tailwind CSS (served by FastAPI)
- **LLM Clients**: Direct API calls using official SDKs
- `openai` for GPT-4
- `anthropic` for Claude
- `google-generativeai` for Gemini
- **Data Visualization**: Chart.js for results analysis
- **Deployment**: Local development server
### Project Structure
```
llm-content-eval/
├── app/
│ ├── __init__.py
│ ├── main.py # FastAPI app entry point
│ ├── models.py # SQLAlchemy/Pydantic models
│ ├── database.py # Database connection/setup
│ ├── schemas.py # Pydantic schemas for API
│ ├── llm_clients.py # LLM API integrations
│ ├── generation_service.py # Business logic for generations
│ ├── evaluation_service.py # Business logic for evaluations
│ ├── analysis_service.py # Results analysis
│ └── routers/
│ ├── __init__.py
│ ├── experiments.py # Experiment endpoints
│ ├── generations.py # Generation endpoints
│ ├── evaluations.py # Evaluation endpoints
│ └── analysis.py # Analysis endpoints
├── static/
│ ├── css/
│ │ └── style.css
│ └── js/
│ └── main.js
├── templates/
│ ├── base.html
│ ├── index.html # Dashboard
│ ├── setup.html # Experiment setup
│ ├── generate.html # Generation progress
│ ├── evaluate.html # Blind evaluation interface
│ └── results.html # Results & analysis
├── data/
│ ├── tasks.json # Task definitions
│ ├── database.db # SQLite database
│ └── baseline_samples/ # User's writing samples
├── config.py # Configuration settings
├── requirements.txt
└── .env # API keys
```
## Data Models
### SQLAlchemy Models (models.py)
```python
from sqlalchemy import Column, Integer, String, Float, DateTime, ForeignKey, JSON, Text, Enum
from sqlalchemy.ext.declarative import declarative_base
from datetime import datetime
import enum
Base = declarative_base()
class ModelProvider(str, enum.Enum):
OPENAI = "openai"
ANTHROPIC = "anthropic"
GOOGLE = "google"
class ContentType(str, enum.Enum):
BLOG_INTRO = "blog_intro"
LINKEDIN = "linkedin"
ANNOUNCEMENT = "announcement"
class PromptStrategy(str, enum.Enum):
STRUCTURED = "structured"
EXAMPLE_BASED = "example_based"
class Experiment(Base):
__tablename__ = "experiments"
id = Column(Integer, primary_key=True)
name = Column(String, nullable=False)
description = Column(Text)
baseline_samples = Column(JSON) # List of sample texts
created_at = Column(DateTime, default=datetime.utcnow)
status = Column(String, default="setup") # setup, generating, evaluating, complete
class Task(Base):
__tablename__ = "tasks"
id = Column(String, primary_key=True) # A, B, C, D, E, F
content_type = Column(Enum(ContentType))
title = Column(String)
description = Column(Text)
structured_prompt = Column(Text)
example_prompt_template = Column(Text)
class Generation(Base):
__tablename__ = "generations"
id = Column(Integer, primary_key=True)
experiment_id = Column(Integer, ForeignKey("experiments.id"))
task_id = Column(String, ForeignKey("tasks.id"))
model_provider = Column(Enum(ModelProvider))
model_name = Column(String) # gpt-4, claude-3-opus, gemini-1.5-pro
prompt_strategy = Column(Enum(PromptStrategy))
prompt_used = Column(Text)
generated_content = Column(Text)
generation_params = Column(JSON) # temperature, max_tokens, etc
timestamp = Column(DateTime, default=datetime.utcnow)
latency_ms = Column(Float)
prompt_tokens = Column(Integer)
completion_tokens = Column(Integer)
cost_usd = Column(Float)
class Evaluation(Base):
__tablename__ = "evaluations"
id = Column(Integer, primary_key=True)
generation_id = Column(Integer, ForeignKey("generations.id"))
experiment_id = Column(Integer, ForeignKey("experiments.id"))
blind_id = Column(String) # Random ID for blind evaluation
# Scores (1-5 scale)
voice_match = Column(Integer)
coherence = Column(Integer)
engaging = Column(Integer)
meets_brief = Column(Integer)
overall_quality = Column(Integer)
# Meta evaluation
edit_time_minutes = Column(Integer)
would_publish = Column(String) # yes, no, with_edits
notes = Column(Text)
evaluated_at = Column(DateTime, default=datetime.utcnow)
evaluation_time_seconds = Column(Integer)
```
### Pydantic Schemas (schemas.py)
```python
from pydantic import BaseModel
from typing import List, Dict, Optional
from datetime import datetime
class ExperimentCreate(BaseModel):
name: str
description: Optional[str]
baseline_samples: List[str]
selected_models: List[Dict[str, str]] # [{"provider": "openai", "model": "gpt-4"}]
selected_strategies: List[str]
selected_tasks: List[str]
class GenerationRequest(BaseModel):
experiment_id: int
run_all: bool = False
specific_combination: Optional[Dict] = None
class EvaluationSubmit(BaseModel):
blind_id: str
voice_match: int
coherence: int
engaging: int
meets_brief: int
overall_quality: int
edit_time_minutes: int
would_publish: str
notes: str
class BlindItem(BaseModel):
blind_id: str
content: str
task_title: str
task_description: str
content_type: str
```
## Task Definitions
### tasks.json
```json
{
"tasks": [
{
"id": "A",
"content_type": "blog_intro",
"title": "B2B Storytelling",
"description": "Write an intro for a blog post about why B2B companies struggle with authentic storytelling",
"structured_prompt": "Write a compelling blog post introduction about why B2B companies struggle with authentic storytelling. Requirements:\n- Target audience: B2B marketing leaders and executives\n- Tone: Professional but conversational, thought-provoking\n- Length: 150-200 words\n- Include: A relatable problem statement, a surprising insight or statistic, and a clear thesis\n- Goal: Hook readers to continue reading about storytelling solutions\n- End with: A transition that previews the value they'll get from the full article",
"example_prompt_template": "Here are two examples of my writing style:\n\n{sample1}\n\n{sample2}\n\nAnalyze the style of these examples, noting:\n- Sentence structure and rhythm\n- Vocabulary and terminology choices\n- How ideas flow and connect\n- Opening and closing techniques\n- Overall voice and personality\n\nNow, write a compelling blog post introduction about why B2B companies struggle with authentic storytelling while matching the style, voice, and approach demonstrated in the examples above. Maintain the same level of sophistication and personality."
},
{
"id": "B",
"content_type": "blog_intro",
"title": "Digital Transformation Costs",
"description": "Write an intro for a blog post about the hidden costs of delaying digital transformation",
"structured_prompt": "Write a compelling blog post introduction about the hidden costs of delaying digital transformation. Requirements:\n- Target audience: C-suite executives in traditional industries\n- Tone: Urgent but not alarmist, data-informed\n- Length: 150-200 words\n- Include: A specific example or scenario, mention of competitive pressure, and opportunity cost\n- Goal: Create urgency without fear-mongering\n- End with: A promise of actionable insights in the article",
"example_prompt_template": "Here are two examples of my writing style:\n\n{sample1}\n\n{sample2}\n\nAnalyze the style of these examples, noting:\n- Sentence structure and rhythm\n- Vocabulary and terminology choices\n- How ideas flow and connect\n- Opening and closing techniques\n- Overall voice and personality\n\nNow, write a compelling blog post introduction about the hidden costs of delaying digital transformation while matching the style, voice, and approach demonstrated in the examples above. Maintain the same level of sophistication and personality."
},
{
"id": "C",
"content_type": "linkedin",
"title": "Failure Lesson",
"description": "Create a LinkedIn post about a lesson learned from a failed project",
"structured_prompt": "Create a LinkedIn post about a lesson learned from a failed project or initiative. Requirements:\n- Target audience: Professional network, industry peers\n- Tone: Vulnerable but professional, educational\n- Length: 200-250 words\n- Include: Brief context of the failure, specific lesson learned, how you apply it now\n- Structure: Hook opening, story, lesson, call for engagement\n- Goal: Build thought leadership through authenticity\n- End with: A question to encourage comments",
"example_prompt_template": "Here are two examples of my writing style:\n\n{sample1}\n\n{sample2}\n\nAnalyze the style of these examples, noting:\n- Sentence structure and rhythm\n- Vocabulary and terminology choices\n- How ideas flow and connect\n- Opening and closing techniques\n- Overall voice and personality\n\nNow, create a LinkedIn post about a lesson learned from a failed project or initiative while matching the style, voice, and approach demonstrated in the examples above. Maintain the same level of sophistication and personality."
},
{
"id": "D",
"content_type": "linkedin",
"title": "Industry Trend",
"description": "Create a LinkedIn post about a new industry trend or insight",
"structured_prompt": "Create a LinkedIn post announcing a new industry trend or insight you've discovered. Requirements:\n- Target audience: Industry professionals and decision-makers\n- Tone: Confident, insightful, accessible\n- Length: 200-250 words\n- Include: The trend/insight, supporting evidence or example, implications for the industry\n- Structure: Bold claim opening, evidence, what this means, action item\n- Goal: Position as thought leader and generate discussion\n- End with: Call-to-action or thought-provoking question",
"example_prompt_template": "Here are two examples of my writing style:\n\n{sample1}\n\n{sample2}\n\nAnalyze the style of these examples, noting:\n- Sentence structure and rhythm\n- Vocabulary and terminology choices\n- How ideas flow and connect\n- Opening and closing techniques\n- Overall voice and personality\n\nNow, create a LinkedIn post announcing a new industry trend or insight you've discovered while matching the style, voice, and approach demonstrated in the examples above. Maintain the same level of sophistication and personality."
},
{
"id": "E",
"content_type": "announcement",
"title": "AI Feature Launch",
"description": "Write an announcement for a new AI-powered feature launch",
"structured_prompt": "Write an announcement for a new AI-powered feature or service launch. Requirements:\n- Target audience: Current customers and prospects\n- Tone: Excited but not hyperbolic, benefit-focused\n- Length: 150-200 words\n- Include: The problem it solves, key capabilities, specific use case, availability\n- Structure: Problem/solution opening, key benefits, proof point, next steps\n- Goal: Drive interest and sign-ups/demos\n- Avoid: Technical jargon, overpromising on AI capabilities",
"example_prompt_template": "Here are two examples of my writing style:\n\n{sample1}\n\n{sample2}\n\nAnalyze the style of these examples, noting:\n- Sentence structure and rhythm\n- Vocabulary and terminology choices\n- How ideas flow and connect\n- Opening and closing techniques\n- Overall voice and personality\n\nNow, write an announcement for a new AI-powered feature or service launch while matching the style, voice, and approach demonstrated in the examples above. Maintain the same level of sophistication and personality."
},
{
"id": "F",
"content_type": "announcement",
"title": "Company Milestone",
"description": "Write an announcement for a company milestone or achievement",
"structured_prompt": "Write an announcement for a company milestone or achievement. Requirements:\n- Target audience: Customers, investors, employees, and industry watchers\n- Tone: Proud but humble, forward-looking\n- Length: 150-200 words\n- Include: Specific achievement, what made it possible, what it means for stakeholders\n- Structure: Achievement statement, context/journey, credit to team/customers, vision forward\n- Goal: Build credibility and momentum\n- End with: Forward-looking statement about what's next",
"example_prompt_template": "Here are two examples of my writing style:\n\n{sample1}\n\n{sample2}\n\nAnalyze the style of these examples, noting:\n- Sentence structure and rhythm\n- Vocabulary and terminology choices\n- How ideas flow and connect\n- Opening and closing techniques\n- Overall voice and personality\n\nNow, write an announcement for a company milestone or achievement while matching the style, voice, and approach demonstrated in the examples above. Maintain the same level of sophistication and personality."
}
]
}
```
## API Endpoints
### Experiments
- `POST /api/experiments` - Create new experiment
- `GET /api/experiments` - List all experiments
- `GET /api/experiments/{id}` - Get experiment details
- `PUT /api/experiments/{id}/status` - Update experiment status
### Generations
- `POST /api/generations/start` - Start generation process
- `GET /api/generations/progress/{experiment_id}` - Get generation progress
- `POST /api/generations/single` - Generate single combination
- `GET /api/generations/{experiment_id}` - Get all generations for experiment
### Evaluations
- `GET /api/evaluations/next/{experiment_id}` - Get next blind item to evaluate
- `POST /api/evaluations` - Submit evaluation
- `GET /api/evaluations/progress/{experiment_id}` - Get evaluation progress
- `GET /api/evaluations/{experiment_id}` - Get all evaluations
### Analysis
- `GET /api/analysis/{experiment_id}/summary` - Get summary statistics
- `GET /api/analysis/{experiment_id}/by-model` - Results grouped by model
- `GET /api/analysis/{experiment_id}/by-strategy` - Results grouped by strategy
- `GET /api/analysis/{experiment_id}/by-task` - Results grouped by task
- `GET /api/analysis/{experiment_id}/export` - Export results as CSV
## LLM Client Implementation (llm_clients.py)
```python
import os
from typing import Dict, Tuple
import time
import openai
import anthropic
import google.generativeai as genai
from config import PRICING
class LLMClient:
def __init__(self):
self.openai_client = openai.OpenAI(api_key=os.getenv("OPENAI_API_KEY"))
self.anthropic_client = anthropic.Anthropic(api_key=os.getenv("ANTHROPIC_API_KEY"))
genai.configure(api_key=os.getenv("GOOGLE_API_KEY"))
def generate(self,
provider: str,
model: str,
prompt: str,
params: Dict) -> Tuple[str, Dict]:
"""
Returns: (generated_content, metadata)
metadata includes: latency_ms, prompt_tokens, completion_tokens, cost_usd
"""
start_time = time.time()
if provider == "openai":
response = self.openai_client.chat.completions.create(
model=model,
messages=[{"role": "user", "content": prompt}],
temperature=params.get("temperature", 0.7),
max_tokens=params.get("max_tokens", 500)
)
content = response.choices[0].message.content
metadata = {
"latency_ms": (time.time() - start_time) * 1000,
"prompt_tokens": response.usage.prompt_tokens,
"completion_tokens": response.usage.completion_tokens,
"cost_usd": self._calculate_cost(provider, model,
response.usage.prompt_tokens,
response.usage.completion_tokens)
}
elif provider == "anthropic":
response = self.anthropic_client.messages.create(
model=model,
messages=[{"role": "user", "content": prompt}],
temperature=params.get("temperature", 0.7),
max_tokens=params.get("max_tokens", 500)
)
content = response.content[0].text
metadata = {
"latency_ms": (time.time() - start_time) * 1000,
"prompt_tokens": response.usage.input_tokens,
"completion_tokens": response.usage.output_tokens,
"cost_usd": self._calculate_cost(provider, model,
response.usage.input_tokens,
response.usage.output_tokens)
}
elif provider == "google":
model_obj = genai.GenerativeModel(model)
response = model_obj.generate_content(
prompt,
generation_config=genai.GenerationConfig(
temperature=params.get("temperature", 0.7),
max_output_tokens=params.get("max_tokens", 500)
)
)
content = response.text
# Note: Token counting for Gemini requires additional setup
metadata = {
"latency_ms": (time.time() - start_time) * 1000,
"prompt_tokens": 0, # Would need token counting
"completion_tokens": 0,
"cost_usd": 0 # Calculate based on Gemini pricing
}
return content, metadata
def _calculate_cost(self, provider, model, prompt_tokens, completion_tokens):
# Implement cost calculation based on current pricing
# Store pricing in config.py
return 0.0 # Placeholder
```
## UI Templates
### Evaluation Interface (evaluate.html)
```html
<!DOCTYPE html>
<html>
<head>
<title>Blind Evaluation</title>
<script src="https://cdn.tailwindcss.com"></script>
</head>
<body class="bg-gray-100">
<div class="container mx-auto px-4 py-8 max-w-4xl">
<h1 class="text-3xl font-bold mb-8">Blind Content Evaluation</h1>
<div class="bg-white rounded-lg shadow-md p-6 mb-6">
<div class="mb-4">
<span class="text-sm text-gray-500">Task:</span>
<h2 class="text-xl font-semibold" id="task-title"></h2>
<p class="text-gray-600" id="task-description"></p>
</div>
<div class="border-t pt-4">
<h3 class="font-semibold mb-2">Generated Content:</h3>
<div class="prose max-w-none" id="content"></div>
</div>
</div>
<form id="evaluation-form" class="bg-white rounded-lg shadow-md p-6">
<h3 class="text-lg font-semibold mb-4">Evaluation Scores (1-5 scale)</h3>
<div class="grid grid-cols-2 gap-4 mb-4">
<div>
<label class="block text-sm font-medium mb-1">Voice/Style Match</label>
<select name="voice_match" class="w-full border rounded px-3 py-2">
<option value="1">1 - Not at all</option>
<option value="2">2 - Slightly</option>
<option value="3">3 - Somewhat</option>
<option value="4">4 - Mostly</option>
<option value="5">5 - Perfect match</option>
</select>
</div>
<div>
<label class="block text-sm font-medium mb-1">Coherence & Flow</label>
<select name="coherence" class="w-full border rounded px-3 py-2">
<option value="1">1 - Poor</option>
<option value="2">2 - Below average</option>
<option value="3">3 - Average</option>
<option value="4">4 - Good</option>
<option value="5">5 - Excellent</option>
</select>
</div>
<div>
<label class="block text-sm font-medium mb-1">Engaging/Compelling</label>
<select name="engaging" class="w-full border rounded px-3 py-2">
<option value="1">1 - Not engaging</option>
<option value="2">2 - Slightly engaging</option>
<option value="3">3 - Moderately engaging</option>
<option value="4">4 - Very engaging</option>
<option value="5">5 - Extremely engaging</option>
</select>
</div>
<div>
<label class="block text-sm font-medium mb-1">Meets Brief Requirements</label>
<select name="meets_brief" class="w-full border rounded px-3 py-2">
<option value="1">1 - Misses completely</option>
<option value="2">2 - Misses key points</option>
<option value="3">3 - Partially meets</option>
<option value="4">4 - Mostly meets</option>
<option value="5">5 - Fully meets</option>
</select>
</div>
</div>
<div class="mb-4">
<label class="block text-sm font-medium mb-1">Overall Quality</label>
<select name="overall_quality" class="w-full border rounded px-3 py-2">
<option value="1">1 - Poor</option>
<option value="2">2 - Below average</option>
<option value="3">3 - Average</option>
<option value="4">4 - Good</option>
<option value="5">5 - Excellent</option>
</select>
</div>
<div class="grid grid-cols-2 gap-4 mb-4">
<div>
<label class="block text-sm font-medium mb-1">Estimated Edit Time (minutes)</label>
<input type="number" name="edit_time_minutes" min="0" class="w-full border rounded px-3 py-2">
</div>
<div>
<label class="block text-sm font-medium mb-1">Would Publish?</label>
<select name="would_publish" class="w-full border rounded px-3 py-2">
<option value="yes">Yes, as is</option>
<option value="with_edits">Yes, with edits</option>
<option value="no">No</option>
</select>
</div>
</div>
<div class="mb-4">
<label class="block text-sm font-medium mb-1">Notes</label>
<textarea name="notes" rows="3" class="w-full border rounded px-3 py-2"></textarea>
</div>
<div class="flex justify-between">
<button type="button" id="skip-btn" class="px-4 py-2 bg-gray-300 rounded hover:bg-gray-400">
Skip for Now
</button>
<button type="submit" class="px-4 py-2 bg-blue-500 text-white rounded hover:bg-blue-600">
Submit & Next
</button>
</div>
</form>
<div class="mt-6 text-center text-sm text-gray-500">
Progress: <span id="progress">0</span> / <span id="total">0</span> evaluated
</div>
</div>
</body>
</html>
```
## Configuration (config.py)
```python
import os
from dotenv import load_dotenv
load_dotenv()
# API Keys
OPENAI_API_KEY = os.getenv("OPENAI_API_KEY")
ANTHROPIC_API_KEY = os.getenv("ANTHROPIC_API_KEY")
GOOGLE_API_KEY = os.getenv("GOOGLE_API_KEY")
# Model configurations
MODELS = {
"openai": {
"model": "gpt-4-turbo-preview",
"params": {
"temperature": 0.7,
"max_tokens": 500
}
},
"anthropic": {
"model": "claude-3-opus-20240229",
"params": {
"temperature": 0.7,
"max_tokens": 500
}
},
"google": {
"model": "gemini-1.5-pro",
"params": {
"temperature": 0.7,
"max_tokens": 500
}
}
}
# Pricing per 1K tokens (update with current pricing)
PRICING = {
"openai": {
"gpt-4-turbo-preview": {
"input": 0.01,
"output": 0.03
}
},
"anthropic": {
"claude-3-opus-20240229": {
"input": 0.015,
"output": 0.075
}
},
"google": {
"gemini-1.5-pro": {
"input": 0.00125,
"output": 0.005
}
}
}
# Database
DATABASE_URL = "sqlite:///data/database.db"
```
## Key Implementation Notes
1. **Blind Evaluation System**:
- Generate random IDs for each piece of content
- Present in randomized order
- Don't reveal source until analysis phase
2. **Generation Order**:
- Randomize the order of generations to avoid API rate limit clustering
- Add small delays between API calls if needed
- Track progress and allow pause/resume
3. **Cost Tracking**:
- Calculate and display running costs during generation
- Show cost per model/strategy in analysis
4. **Analysis Features**:
- Heatmap of scores (model vs strategy)
- Best/worst performing combinations
- Statistical significance tests (if you add them later)
- Export all data to CSV for external analysis
5. **Error Handling**:
- Retry logic for API failures
- Save partial progress
- Clear error messages to user
6. **User Flow**:
1. Setup: Upload baseline samples, select models/tasks
2. Generate: Run all 36 generations with progress tracking
3. Evaluate: Blind evaluation interface
4. Analyze: View results and identify best performers
## Getting Started Instructions
1. Install dependencies:
```bash
pip install fastapi uvicorn sqlalchemy openai anthropic google-generativeai python-dotenv
```
2. Set up environment variables in `.env`:
```
OPENAI_API_KEY=your_key_here
ANTHROPIC_API_KEY=your_key_here
GOOGLE_API_KEY=your_key_here
```
3. Initialize database and load tasks:
```bash
python -c "from app.database import init_db; init_db()"
python -c "from app.database import load_tasks; load_tasks('data/tasks.json')"
```
4. Run the application:
```bash
uvicorn app.main:app --reload
```
5. Access at `http://localhost:8000`
This specification provides everything needed to build the complete evaluation system. The app will allow systematic testing of different LLMs and prompting strategies, with blind evaluation to ensure unbiased results and comprehensive analysis to identify the best approach for each content type.