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#!/usr/bin/env python3
"""Test script for chained analysis approach.
This script demonstrates breaking the monolithic analysis prompt into
focused, sequential steps that build on each other.
The chain:
1. Layout Detection (visual only) - What's the layout of each page?
2. Document Type (visual + layout) - What kind of document is this?
3. Heading Structure (needs layout for reading order) - Build heading tree
4. Page Features (per-page detection) - Images, tables, code, math
5. Agent Routing (logic over data) - Which agents are needed?
Run from project root:
python scripts/test_chained_analysis.py [pdf_path] [--pages N]
Examples:
# Test with attention paper (two-column)
python scripts/test_chained_analysis.py project-docs/pdfs/07_attention_transformer_paper.pdf
# Test first 3 pages only
python scripts/test_chained_analysis.py project-docs/pdfs/07_attention_transformer_paper.pdf --pages 3
"""
import asyncio
import sys
import argparse
import json
import time
from pathlib import Path
from typing import Literal
from pydantic import BaseModel, Field
from pydantic_ai import Agent, BinaryContent
from pydantic_ai.models.bedrock import BedrockConverseModel
# Add src to path for imports
sys.path.insert(0, str(Path(__file__).parent.parent))
from src.services.pdf_converter import PDFConverter, PageData
from src.agents.model_tiers import MODEL_TIER_MAP, ModelTier
# =============================================================================
# Step 1: Layout Detection
# =============================================================================
class PageLayout(BaseModel):
"""Layout for a single page."""
page_num: int
layout_type: Literal["single_column", "two_column", "mixed"]
confidence: float = Field(ge=0.0, le=1.0)
evidence: str = Field(description="One sentence describing visual evidence")
class LayoutOutput(BaseModel):
"""Output from layout detection step."""
pages: list[PageLayout]
LAYOUT_SYSTEM_PROMPT = """You analyze PDF page layouts. For each page image, determine the layout type.
Layout types:
- single_column: Text flows across the full page width in one column
- two_column: Page is divided into two columns with a vertical gutter between them
- mixed: Single page has both single-column and two-column sections
Visual cues for two_column:
- Vertical whitespace (gutter) dividing page roughly in half
- Text blocks that don't span full page width
- Parallel content at the same vertical position on left and right
Be concise. State your visual evidence in one sentence."""
# =============================================================================
# Step 2: Document Type Classification
# =============================================================================
class DocumentTypeOutput(BaseModel):
"""Output from document type classification."""
document_type: Literal["syllabus", "lecture_notes", "exam", "research_paper", "handout", "thesis", "report", "other"]
confidence: float = Field(ge=0.0, le=1.0)
evidence: str = Field(description="Key visual indicators for this classification")
DOCTYPE_SYSTEM_PROMPT = """You classify academic document types based on visual appearance.
Document types:
- syllabus: Course outline with schedule, grading, policies
- lecture_notes: Class slides or notes
- exam: Test, quiz, or assignment with questions
- research_paper: Academic paper with abstract, sections, citations (often two-column)
- handout: Supplementary material, worksheets
- thesis: Long-form academic work with chapters
- report: Technical or organizational report
- other: Doesn't fit above categories
Look at the first page primarily. Consider:
- Title format and positioning
- Presence of abstract
- Two-column layout (suggests research paper)
- Course codes or instructor names (suggests syllabus)
- Question numbering (suggests exam)
Be concise."""
# =============================================================================
# Step 3: Heading Structure
# =============================================================================
class Heading(BaseModel):
"""A heading in the document."""
level: int = Field(ge=1, le=6, description="Heading level 1-6 (H1-H6)")
text: str = Field(description="Exact heading text")
page_num: int
section_number: str | None = Field(default=None, description="Section number if present (e.g., '1.1', '2')")
class HeadingsOutput(BaseModel):
"""Output from heading structure extraction."""
headings: list[Heading]
reading_order_notes: str = Field(description="Notes about reading order decisions for multi-column pages")
HEADINGS_SYSTEM_PROMPT = """You extract document heading structure from PDF pages.
For each heading you find, record:
- level: H1 (main title) through H6 (deepest subsection)
- text: Exact heading text as it appears
- page_num: Which page it's on
- section_number: If the heading has a number like "1.1" or "Chapter 2"
CRITICAL for reading order:
- For single_column pages: read top to bottom
- For two_column pages: read LEFT column completely (top to bottom), THEN right column (top to bottom)
- This means a heading in the left column comes BEFORE a heading at the same height in the right column
Verify parent-child relationships:
- Section 5.1 must come AFTER section 5
- Don't skip levels (H1 → H3 without H2) unless that's genuinely how the document is structured
Be thorough but concise."""
# =============================================================================
# Step 4: Page Features
# =============================================================================
class PageFeatures(BaseModel):
"""Features detected on a single page."""
page_num: int
has_images: bool = Field(description="Informative images (charts, diagrams, photos) - NOT decorative")
image_count: int = Field(ge=0)
has_tables: bool = Field(description="Data tables with rows/columns")
table_count: int = Field(ge=0)
has_lists: bool = Field(description="Bulleted or numbered lists")
has_code_blocks: bool = Field(description="Programming code or command-line content")
has_math: bool = Field(description="Mathematical equations or formulas")
has_figures_with_captions: bool = Field(description="Figures that have caption text")
class FeaturesOutput(BaseModel):
"""Output from page features detection."""
pages: list[PageFeatures]
FEATURES_SYSTEM_PROMPT = """You detect content features on PDF pages.
For each page, identify:
- Informative images: Charts, graphs, diagrams, photos, screenshots (NOT decorative borders/logos)
- Data tables: Structured rows and columns of data
- Lists: Bulleted or numbered items
- Code blocks: Programming code, terminal output, monospace technical content
- Math: Equations, formulas, mathematical notation
- Figures with captions: Images that have associated caption text
Be accurate with counts. Only count informative content, not decorative elements."""
# =============================================================================
# Step 5: Agent Routing
# =============================================================================
class RoutingOutput(BaseModel):
"""Output from agent routing decision."""
required_agents: list[Literal["figures", "tables", "structure", "typography"]]
reasoning: str = Field(description="Why each agent is or isn't needed")
ROUTING_SYSTEM_PROMPT = """You decide which specialized agents should process a document.
Available agents:
- figures: Analyzes images, generates alt text. Needed if document has informative images.
- tables: Analyzes table structure for accessibility. Needed if document has data tables.
- structure: Validates heading hierarchy and reading order. Needed for complex hierarchies or two-column layouts.
- typography: Checks if visual styling conveys meaning (color-coding, bold for emphasis). Needed if styling is semantic.
Rules:
- Be conservative - only include agents that are clearly needed
- If document has two-column layout, include "structure" (reading order verification)
- If document has images, include "figures"
- If document has tables, include "tables"
- Only include "typography" if you see color-coded content or meaningful bold/italic usage"""
# =============================================================================
# Main Test Logic
# =============================================================================
async def run_chained_analysis(
pages: list[PageData],
max_pages: int | None = None,
use_sonnet: bool = False,
parallel: bool = False,
) -> dict:
"""Run the chained analysis on PDF pages.
Args:
pages: List of PageData from PDF conversion
max_pages: Limit analysis to first N pages (for faster testing)
use_sonnet: Use Sonnet instead of Haiku (for comparison)
parallel: Run steps 2-4 in parallel after step 1
Returns:
Dictionary with results from each step
"""
# Limit pages if requested
if max_pages:
pages = pages[:max_pages]
total_pages = len(pages)
print(f"\n{'='*60}")
print(f"Running chained analysis on {total_pages} pages")
print(f"Model: {'Sonnet' if use_sonnet else 'Haiku'}")
print(f"Mode: {'PARALLEL (steps 2-4 concurrent)' if parallel else 'SEQUENTIAL'}")
print(f"{'='*60}\n")
# Select model
model_tier = ModelTier.REASONING if use_sonnet else ModelTier.EFFICIENT
model = BedrockConverseModel(MODEL_TIER_MAP[model_tier])
# Prepare page images for multimodal input
page_images = [
BinaryContent(data=bytes.fromhex(page.image_base64) if len(page.image_base64) < 100
else __import__('base64').b64decode(page.image_base64),
media_type="image/png")
for page in pages
]
results = {}
total_tokens = {"input": 0, "output": 0}
# -------------------------------------------------------------------------
# Step 1: Layout Detection
# -------------------------------------------------------------------------
print("Step 1: Layout Detection")
print("-" * 40)
start = time.time()
layout_agent = Agent(
model,
output_type=LayoutOutput,
system_prompt=LAYOUT_SYSTEM_PROMPT,
)
layout_prompt = [
f"Analyze the layout of these {total_pages} PDF pages. For each page, determine if it's single_column, two_column, or mixed.",
*page_images
]
layout_result = await layout_agent.run(layout_prompt)
layout_output = layout_result.output
elapsed = time.time() - start
usage = layout_result.usage()
total_tokens["input"] += usage.input_tokens or 0
total_tokens["output"] += usage.output_tokens or 0
print(f" Time: {elapsed:.1f}s | Tokens: {usage.input_tokens} in, {usage.output_tokens} out")
for pl in layout_output.pages:
print(f" Page {pl.page_num}: {pl.layout_type} (conf: {pl.confidence:.2f})")
print(f" Evidence: {pl.evidence}")
results["layout"] = layout_output.model_dump()
# Summarize layouts for next steps
layout_summary = ", ".join([
f"Page {p.page_num}: {p.layout_type}"
for p in layout_output.pages
])
has_two_column = any(p.layout_type == "two_column" for p in layout_output.pages)
# -------------------------------------------------------------------------
# Steps 2-4: Run in parallel or sequential based on flag
# -------------------------------------------------------------------------
if parallel:
# PARALLEL MODE: Steps 2, 3, 4 run concurrently
print("\n[PARALLEL] Running Steps 2, 3, 4 concurrently...")
print("-" * 40)
parallel_start = time.time()
async def run_doctype():
"""Step 2: Document Type Classification"""
agent = Agent(model, output_type=DocumentTypeOutput, system_prompt=DOCTYPE_SYSTEM_PROMPT)
prompt = [f"What type of document is this? Layout info: {layout_summary}", page_images[0]]
if len(page_images) > 1:
prompt.append(page_images[-1])
return await agent.run(prompt)
async def run_headings():
"""Step 3: Heading Structure (uses layout info)"""
agent = Agent(model, output_type=HeadingsOutput, system_prompt=HEADINGS_SYSTEM_PROMPT)
reading_order_guidance = ""
if has_two_column:
two_col_pages = [p.page_num for p in layout_output.pages if p.layout_type == "two_column"]
reading_order_guidance = f"\n\nCRITICAL: Pages {two_col_pages} are two-column. Read each column top-to-bottom, left column first, then right column."
prompt = [f"Extract all headings from this document.{reading_order_guidance}", *page_images]
return await agent.run(prompt)
async def run_features():
"""Step 4: Page Features Detection"""
agent = Agent(model, output_type=FeaturesOutput, system_prompt=FEATURES_SYSTEM_PROMPT)
prompt = [f"Detect content features on each of these {total_pages} pages.", *page_images]
return await agent.run(prompt)
# Run all three in parallel
doctype_result, headings_result, features_result = await asyncio.gather(
run_doctype(),
run_headings(),
run_features(),
)
parallel_elapsed = time.time() - parallel_start
# Extract outputs
doctype_output = doctype_result.output
headings_output = headings_result.output
features_output = features_result.output
# Aggregate tokens
for result in [doctype_result, headings_result, features_result]:
usage = result.usage()
total_tokens["input"] += usage.input_tokens or 0
total_tokens["output"] += usage.output_tokens or 0
# Print results
print(f"\n Step 2 (Doc Type): {doctype_output.document_type} (conf: {doctype_output.confidence:.2f})")
print(f" Evidence: {doctype_output.evidence}")
print(f"\n Step 3 (Headings): Found {len(headings_output.headings)} headings")
for h in headings_output.headings[:5]:
prefix = f"[{h.section_number}] " if h.section_number else ""
print(f" H{h.level} (p{h.page_num}): {prefix}{h.text[:40]}{'...' if len(h.text) > 40 else ''}")
if len(headings_output.headings) > 5:
print(f" ... and {len(headings_output.headings) - 5} more")
total_images = sum(p.image_count for p in features_output.pages)
total_tables = sum(p.table_count for p in features_output.pages)
pages_with_code = sum(1 for p in features_output.pages if p.has_code_blocks)
pages_with_math = sum(1 for p in features_output.pages if p.has_math)
print(f"\n Step 4 (Features): {total_images} images, {total_tables} tables, {pages_with_code} code pages, {pages_with_math} math pages")
print(f"\n [PARALLEL] Total time for steps 2-4: {parallel_elapsed:.1f}s")
results["document_type"] = doctype_output.model_dump()
results["headings"] = headings_output.model_dump()
results["features"] = features_output.model_dump()
else:
# SEQUENTIAL MODE: Original step-by-step execution
# -------------------------------------------------------------------------
# Step 2: Document Type Classification
# -------------------------------------------------------------------------
print("\nStep 2: Document Type Classification")
print("-" * 40)
start = time.time()
doctype_agent = Agent(
model,
output_type=DocumentTypeOutput,
system_prompt=DOCTYPE_SYSTEM_PROMPT,
)
# Only need first page (and maybe last) for document type
doctype_prompt = [
f"What type of document is this? Layout info: {layout_summary}",
page_images[0], # First page usually enough
]
if len(page_images) > 1:
doctype_prompt.append(page_images[-1]) # Add last page for context
doctype_result = await doctype_agent.run(doctype_prompt)
doctype_output = doctype_result.output
elapsed = time.time() - start
usage = doctype_result.usage()
total_tokens["input"] += usage.input_tokens or 0
total_tokens["output"] += usage.output_tokens or 0
print(f" Time: {elapsed:.1f}s | Tokens: {usage.input_tokens} in, {usage.output_tokens} out")
print(f" Type: {doctype_output.document_type} (conf: {doctype_output.confidence:.2f})")
print(f" Evidence: {doctype_output.evidence}")
results["document_type"] = doctype_output.model_dump()
# -------------------------------------------------------------------------
# Step 3: Heading Structure
# -------------------------------------------------------------------------
print("\nStep 3: Heading Structure Extraction")
print("-" * 40)
start = time.time()
headings_agent = Agent(
model,
output_type=HeadingsOutput,
system_prompt=HEADINGS_SYSTEM_PROMPT,
)
# Include layout info in prompt for correct reading order
reading_order_guidance = ""
if has_two_column:
two_col_pages = [p.page_num for p in layout_output.pages if p.layout_type == "two_column"]
reading_order_guidance = f"\n\nCRITICAL: Pages {two_col_pages} are two-column. Read each column top-to-bottom, left column first, then right column."
headings_prompt = [
f"Extract all headings from this {doctype_output.document_type} document.{reading_order_guidance}",
*page_images
]
headings_result = await headings_agent.run(headings_prompt)
headings_output = headings_result.output
elapsed = time.time() - start
usage = headings_result.usage()
total_tokens["input"] += usage.input_tokens or 0
total_tokens["output"] += usage.output_tokens or 0
print(f" Time: {elapsed:.1f}s | Tokens: {usage.input_tokens} in, {usage.output_tokens} out")
print(f" Found {len(headings_output.headings)} headings:")
for h in headings_output.headings[:10]: # Show first 10
prefix = f"[{h.section_number}] " if h.section_number else ""
print(f" H{h.level} (p{h.page_num}): {prefix}{h.text[:50]}{'...' if len(h.text) > 50 else ''}")
if len(headings_output.headings) > 10:
print(f" ... and {len(headings_output.headings) - 10} more")
print(f" Reading order notes: {headings_output.reading_order_notes}")
results["headings"] = headings_output.model_dump()
# -------------------------------------------------------------------------
# Step 4: Page Features Detection
# -------------------------------------------------------------------------
print("\nStep 4: Page Features Detection")
print("-" * 40)
start = time.time()
features_agent = Agent(
model,
output_type=FeaturesOutput,
system_prompt=FEATURES_SYSTEM_PROMPT,
)
features_prompt = [
f"Detect content features on each of these {total_pages} pages.",
*page_images
]
features_result = await features_agent.run(features_prompt)
features_output = features_result.output
elapsed = time.time() - start
usage = features_result.usage()
total_tokens["input"] += usage.input_tokens or 0
total_tokens["output"] += usage.output_tokens or 0
print(f" Time: {elapsed:.1f}s | Tokens: {usage.input_tokens} in, {usage.output_tokens} out")
total_images = sum(p.image_count for p in features_output.pages)
total_tables = sum(p.table_count for p in features_output.pages)
pages_with_code = sum(1 for p in features_output.pages if p.has_code_blocks)
pages_with_math = sum(1 for p in features_output.pages if p.has_math)
print(f" Summary: {total_images} images, {total_tables} tables, {pages_with_code} pages with code, {pages_with_math} pages with math")
for pf in features_output.pages:
features_list = []
if pf.has_images: features_list.append(f"{pf.image_count} img")
if pf.has_tables: features_list.append(f"{pf.table_count} tbl")
if pf.has_lists: features_list.append("lists")
if pf.has_code_blocks: features_list.append("code")
if pf.has_math: features_list.append("math")
if features_list:
print(f" Page {pf.page_num}: {', '.join(features_list)}")
results["features"] = features_output.model_dump()
# -------------------------------------------------------------------------
# Step 5: Agent Routing
# -------------------------------------------------------------------------
print("\nStep 5: Agent Routing Decision")
print("-" * 40)
start = time.time()
routing_agent = Agent(
model,
output_type=RoutingOutput,
system_prompt=ROUTING_SYSTEM_PROMPT,
)
# This step doesn't need images - pure logic over collected data
routing_prompt = f"""Based on the analysis so far, decide which specialized agents are needed.
Document type: {doctype_output.document_type}
Layout: {layout_summary}
Has two-column pages: {has_two_column}
Total images: {total_images}
Total tables: {total_tables}
Heading depth: {max((h.level for h in headings_output.headings), default=1)}
Pages with code: {pages_with_code}
Pages with math: {pages_with_math}
Which agents should process this document?"""
routing_result = await routing_agent.run(routing_prompt)
routing_output = routing_result.output
elapsed = time.time() - start
usage = routing_result.usage()
total_tokens["input"] += usage.input_tokens or 0
total_tokens["output"] += usage.output_tokens or 0
print(f" Time: {elapsed:.1f}s | Tokens: {usage.input_tokens} in, {usage.output_tokens} out")
print(f" Required agents: {routing_output.required_agents}")
print(f" Reasoning: {routing_output.reasoning}")
results["routing"] = routing_output.model_dump()
# -------------------------------------------------------------------------
# Summary
# -------------------------------------------------------------------------
print(f"\n{'='*60}")
print("ANALYSIS COMPLETE")
print(f"{'='*60}")
print(f"Total tokens: {total_tokens['input']} input, {total_tokens['output']} output")
print(f"\nDocument: {doctype_output.document_type}")
print(f"Layout: {'Two-column' if has_two_column else 'Single-column'}")
print(f"Headings: {len(headings_output.headings)}")
print(f"Images: {total_images}, Tables: {total_tables}")
print(f"Agents needed: {routing_output.required_agents}")
return results
async def main():
parser = argparse.ArgumentParser(description="Test chained analysis approach")
parser.add_argument("pdf_path", help="Path to PDF file to analyze")
parser.add_argument("--pages", type=int, help="Limit to first N pages")
parser.add_argument("--sonnet", action="store_true", help="Use Sonnet instead of Haiku")
parser.add_argument("--parallel", action="store_true", help="Run steps 2-4 in parallel")
parser.add_argument("--output", "-o", help="Save results to JSON file")
args = parser.parse_args()
pdf_path = Path(args.pdf_path)
if not pdf_path.exists():
print(f"Error: PDF not found: {pdf_path}")
sys.exit(1)
print(f"Loading PDF: {pdf_path}")
# Convert PDF to page images
converter = PDFConverter()
with open(pdf_path, "rb") as f:
pdf_content = f.read()
print(f"Converting PDF ({len(pdf_content)} bytes)...")
conversion_result = await converter.convert_with_page_images(pdf_content)
print(f"Converted: {conversion_result.total_pages} pages")
# Run chained analysis
results = await run_chained_analysis(
pages=conversion_result.pages,
max_pages=args.pages,
use_sonnet=args.sonnet,
parallel=args.parallel,
)
# Save results if requested
if args.output:
output_path = Path(args.output)
with open(output_path, "w") as f:
json.dump(results, f, indent=2)
print(f"\nResults saved to: {output_path}")
if __name__ == "__main__":
asyncio.run(main())