#!/usr/bin/env python3 """ Build the 8-slide Retrieval Benchmark Rerun PDF for leadership. Uses matplotlib for slide rendering with a dark professional theme. Embeds figures from figures/ directory. """ import matplotlib matplotlib.use("Agg") import matplotlib.pyplot as plt import matplotlib.image as mpimg from matplotlib.backends.backend_pdf import PdfPages from pathlib import Path import textwrap # ── Paths ────────────────────────────────────────────────────────────── ROOT = Path(__file__).resolve().parent.parent FIG_DIR = ROOT / "figures" OUT_PDF = ROOT / "Retrieval_Benchmark_Rerun_Leadership.pdf" # ── Theme ────────────────────────────────────────────────────────────── BG = "#0d1b2a" # deep navy BG_CARD = "#1b2d45" # slightly lighter card ACCENT = "#00b4d8" # bright cyan accent ACCENT2 = "#48cae4" # lighter cyan GOLD = "#ffd60a" # gold for callout numbers WHITE = "#e0e1dd" MUTED = "#778da9" RED_SOFT = "#ef476f" SLIDE_W, SLIDE_H = 16, 9 # 16:9 def new_slide(): fig = plt.figure(figsize=(SLIDE_W, SLIDE_H), facecolor=BG) fig.subplots_adjust(left=0, right=1, top=1, bottom=0) return fig def add_footer(fig, slide_num, total=8): fig.text(0.96, 0.025, f"{slide_num}/{total}", fontsize=11, color=MUTED, ha="right", va="bottom", fontfamily="sans-serif") fig.text(0.04, 0.025, "Confidential", fontsize=9, color=MUTED, ha="left", va="bottom", fontfamily="sans-serif", style="italic") # ══════════════════════════════════════════════════════════════════════ # SLIDE 1 — Title # ══════════════════════════════════════════════════════════════════════ def slide_title(): fig = new_slide() # Decorative accent bar ax = fig.add_axes([0.06, 0.42, 0.005, 0.22]) ax.set_facecolor(ACCENT) ax.set_xticks([]); ax.set_yticks([]) for spine in ax.spines.values(): spine.set_visible(False) fig.text(0.09, 0.62, "Retrieval Benchmark Rerun", fontsize=48, fontweight="bold", color=WHITE, fontfamily="sans-serif", va="bottom") fig.text(0.09, 0.54, "Updated Results", fontsize=42, fontweight="light", color=ACCENT, fontfamily="sans-serif", va="bottom") fig.text(0.09, 0.40, "April 10, 2026 | Leadership Update", fontsize=20, color=MUTED, fontfamily="sans-serif") fig.text(0.09, 0.33, "doany Retrieval Team", fontsize=18, color=MUTED, fontfamily="sans-serif") # Bottom accent line ax2 = fig.add_axes([0.06, 0.12, 0.35, 0.003]) ax2.set_facecolor(ACCENT) ax2.set_xticks([]); ax2.set_yticks([]) for spine in ax2.spines.values(): spine.set_visible(False) add_footer(fig, 1) return fig # ══════════════════════════════════════════════════════════════════════ # SLIDE 2 — Why We Reran # ══════════════════════════════════════════════════════════════════════ def slide_why(): fig = new_slide() fig.text(0.06, 0.88, "Why We Reran", fontsize=36, fontweight="bold", color=WHITE, fontfamily="sans-serif") # Accent underline ax_line = fig.add_axes([0.06, 0.855, 0.18, 0.003]) ax_line.set_facecolor(ACCENT); ax_line.set_xticks([]); ax_line.set_yticks([]) for s in ax_line.spines.values(): s.set_visible(False) # Bug callout card ax_card = fig.add_axes([0.06, 0.48, 0.40, 0.32]) ax_card.set_facecolor(BG_CARD); ax_card.set_xticks([]); ax_card.set_yticks([]) for s in ax_card.spines.values(): s.set_color(ACCENT); s.set_linewidth(1.5) fig.text(0.08, 0.76, "Chunking Overlap Bug (PR #347)", fontsize=22, fontweight="bold", color=RED_SOFT, fontfamily="sans-serif") fig.text(0.08, 0.68, "Silently dropped ~8% of document chunks\nin the indexing pipeline", fontsize=18, color=WHITE, fontfamily="sans-serif", linespacing=1.5) fig.text(0.08, 0.54, "Disproportionately hurt models trained\nwith overlapping context windows", fontsize=16, color=MUTED, fontfamily="sans-serif", linespacing=1.5) # Key facts — right side facts = [ ("Previous run", "Mar 28, 2026"), ("Rerun", "Apr 10, 2026"), ("Dataset", "doany-skills-eval-v2"), ("Queries", "4,200"), ("Documents", "58,000"), ] y = 0.76 for label, value in facts: fig.text(0.56, y, label, fontsize=16, color=MUTED, fontfamily="sans-serif") fig.text(0.78, y, value, fontsize=16, fontweight="bold", color=WHITE, fontfamily="sans-serif") y -= 0.065 add_footer(fig, 2) return fig # ══════════════════════════════════════════════════════════════════════ # SLIDE 3 — Headline Results (with figure) # ══════════════════════════════════════════════════════════════════════ def slide_headline(): fig = new_slide() fig.text(0.06, 0.88, "Headline Results", fontsize=36, fontweight="bold", color=WHITE, fontfamily="sans-serif") ax_line = fig.add_axes([0.06, 0.855, 0.22, 0.003]) ax_line.set_facecolor(ACCENT); ax_line.set_xticks([]); ax_line.set_yticks([]) for s in ax_line.spines.values(): s.set_visible(False) # Embed figure — left/center img_path = FIG_DIR / "retrieval_accuracy.png" ax_img = fig.add_axes([0.04, 0.08, 0.55, 0.72]) ax_img.set_facecolor(BG) ax_img.set_xticks([]); ax_img.set_yticks([]) for s in ax_img.spines.values(): s.set_visible(False) if img_path.exists(): img = mpimg.imread(str(img_path)) ax_img.imshow(img, aspect="auto") # Callout numbers — right side fig.text(0.66, 0.76, "84.7%", fontsize=56, fontweight="bold", color=GOLD, fontfamily="sans-serif") fig.text(0.66, 0.69, "doany-ret-v3 Top-5 Accuracy", fontsize=16, color=WHITE, fontfamily="sans-serif") fig.text(0.66, 0.55, "+3.5pp", fontsize=44, fontweight="bold", color=ACCENT, fontfamily="sans-serif") fig.text(0.66, 0.49, "vs. previous run (81.2%)", fontsize=15, color=MUTED, fontfamily="sans-serif") fig.text(0.66, 0.36, "5.8pp", fontsize=40, fontweight="bold", color=ACCENT2, fontfamily="sans-serif") fig.text(0.66, 0.30, "lead over BGE-M3 (was 3.1pp)", fontsize=15, color=MUTED, fontfamily="sans-serif") add_footer(fig, 3) return fig # ══════════════════════════════════════════════════════════════════════ # SLIDE 4 — Recall@K Deep Dive # ══════════════════════════════════════════════════════════════════════ def slide_recall(): fig = new_slide() fig.text(0.06, 0.88, "Recall@K Deep Dive", fontsize=36, fontweight="bold", color=WHITE, fontfamily="sans-serif") ax_line = fig.add_axes([0.06, 0.855, 0.24, 0.003]) ax_line.set_facecolor(ACCENT); ax_line.set_xticks([]); ax_line.set_yticks([]) for s in ax_line.spines.values(): s.set_visible(False) # Embed figure — center img_path = FIG_DIR / "recall_at_k.png" ax_img = fig.add_axes([0.04, 0.08, 0.55, 0.72]) ax_img.set_facecolor(BG) ax_img.set_xticks([]); ax_img.set_yticks([]) for s in ax_img.spines.values(): s.set_visible(False) if img_path.exists(): img = mpimg.imread(str(img_path)) ax_img.imshow(img, aspect="auto") # Callout — right fig.text(0.66, 0.74, "92%", fontsize=56, fontweight="bold", color=GOLD, fontfamily="sans-serif") fig.text(0.66, 0.67, "Recall@10 for doany-ret-v3", fontsize=16, color=WHITE, fontfamily="sans-serif") fig.text(0.66, 0.61, "up from 88% (+4pp)", fontsize=15, color=MUTED, fontfamily="sans-serif") # Secondary callout fig.text(0.66, 0.46, "BM25 Unaffected", fontsize=22, fontweight="bold", color=MUTED, fontfamily="sans-serif") fig.text(0.66, 0.40, "Sparse retrieval doesn't depend\non chunk overlap", fontsize=14, color=MUTED, fontfamily="sans-serif", linespacing=1.5) # Mini table — right side models = [ ("doany-ret-v3", "92%", GOLD), ("BGE-M3", "87%", WHITE), ("E5-Large", "86%", WHITE), ("ColBERT v2", "84%", WHITE), ("BM25", "74%", MUTED), ] y = 0.28 fig.text(0.66, y + 0.04, "Recall@10 Leaderboard", fontsize=14, fontweight="bold", color=ACCENT, fontfamily="sans-serif") for name, val, clr in models: fig.text(0.66, y, name, fontsize=13, color=MUTED, fontfamily="sans-serif") fig.text(0.88, y, val, fontsize=13, fontweight="bold", color=clr, fontfamily="sans-serif", ha="right") y -= 0.04 add_footer(fig, 4) return fig # ══════════════════════════════════════════════════════════════════════ # SLIDE 5 — Latency Profile # ══════════════════════════════════════════════════════════════════════ def slide_latency(): fig = new_slide() fig.text(0.06, 0.88, "Latency Profile", fontsize=36, fontweight="bold", color=WHITE, fontfamily="sans-serif") ax_line = fig.add_axes([0.06, 0.855, 0.19, 0.003]) ax_line.set_facecolor(ACCENT); ax_line.set_xticks([]); ax_line.set_yticks([]) for s in ax_line.spines.values(): s.set_visible(False) # Embed figure img_path = FIG_DIR / "latency_comparison.png" ax_img = fig.add_axes([0.04, 0.08, 0.55, 0.72]) ax_img.set_facecolor(BG) ax_img.set_xticks([]); ax_img.set_yticks([]) for s in ax_img.spines.values(): s.set_visible(False) if img_path.exists(): img = mpimg.imread(str(img_path)) ax_img.imshow(img, aspect="auto") # Key message fig.text(0.66, 0.76, "Zero Overhead", fontsize=36, fontweight="bold", color=GOLD, fontfamily="sans-serif") fig.text(0.66, 0.69, "Fix was index-side only\nNo inference latency change", fontsize=16, color=WHITE, fontfamily="sans-serif", linespacing=1.5) # doany-ret-v3 stats fig.text(0.66, 0.52, "doany-ret-v3", fontsize=18, fontweight="bold", color=ACCENT, fontfamily="sans-serif") stats = [("p50", "34 ms"), ("p95", "71 ms")] y = 0.44 for label, val in stats: fig.text(0.66, y, label, fontsize=20, color=MUTED, fontfamily="sans-serif") fig.text(0.78, y, val, fontsize=20, fontweight="bold", color=WHITE, fontfamily="sans-serif") y -= 0.06 fig.text(0.66, 0.27, "Fastest neural retriever", fontsize=16, fontweight="bold", color=ACCENT2, fontfamily="sans-serif") fig.text(0.66, 0.22, "2x faster than E5-Large (p50=62ms)\n1.7x faster than BGE-M3 (p50=58ms)", fontsize=13, color=MUTED, fontfamily="sans-serif", linespacing=1.5) add_footer(fig, 5) return fig # ══════════════════════════════════════════════════════════════════════ # SLIDE 6 — Related Work # ══════════════════════════════════════════════════════════════════════ def slide_related(): fig = new_slide() fig.text(0.06, 0.88, "Related Work & Benchmarks", fontsize=36, fontweight="bold", color=WHITE, fontfamily="sans-serif") ax_line = fig.add_axes([0.06, 0.855, 0.30, 0.003]) ax_line.set_facecolor(ACCENT); ax_line.set_xticks([]); ax_line.set_yticks([]) for s in ax_line.spines.values(): s.set_visible(False) # Two-column layout # Left column: Key models fig.text(0.06, 0.77, "Retrieval Models", fontsize=20, fontweight="bold", color=ACCENT, fontfamily="sans-serif") refs_left = [ ("DPR", "Karpukhin et al., 2020", "Dense Passage Retrieval — pioneered\ndual-encoder neural retrieval"), ("ColBERT", "Khattab & Zaharia, 2020", "Late-interaction retrieval with\ntoken-level matching"), ("E5", "Wang et al., 2022", "Unified text embeddings for\ntransfer learning"), ("BGE-M3", "Chen et al., 2024", "Multilingual embeddings —\nstrongest open-source baseline"), ] y = 0.70 for name, cite, desc in refs_left: fig.text(0.06, y, name, fontsize=16, fontweight="bold", color=WHITE, fontfamily="sans-serif") fig.text(0.18, y, cite, fontsize=13, color=MUTED, fontfamily="sans-serif") fig.text(0.06, y - 0.045, desc, fontsize=11, color=MUTED, fontfamily="sans-serif", linespacing=1.3) y -= 0.13 # Right column: Benchmarks fig.text(0.55, 0.77, "Evaluation Benchmarks", fontsize=20, fontweight="bold", color=ACCENT, fontfamily="sans-serif") fig.text(0.55, 0.69, "MTEB", fontsize=16, fontweight="bold", color=WHITE, fontfamily="sans-serif") fig.text(0.67, 0.69, "Muennighoff et al., 2023", fontsize=13, color=MUTED, fontfamily="sans-serif") fig.text(0.55, 0.645, "Massive Text Embedding Benchmark\n58 datasets across 8 tasks", fontsize=11, color=MUTED, fontfamily="sans-serif", linespacing=1.3) fig.text(0.55, 0.56, "BEIR", fontsize=16, fontweight="bold", color=WHITE, fontfamily="sans-serif") fig.text(0.67, 0.56, "Thakur et al., 2021", fontsize=13, color=MUTED, fontfamily="sans-serif") fig.text(0.55, 0.515, "Heterogeneous retrieval benchmark\n18 diverse retrieval datasets", fontsize=11, color=MUTED, fontfamily="sans-serif", linespacing=1.3) # Comparison callout card ax_card = fig.add_axes([0.55, 0.14, 0.40, 0.28]) ax_card.set_facecolor(BG_CARD); ax_card.set_xticks([]); ax_card.set_yticks([]) for s in ax_card.spines.values(): s.set_color(GOLD); s.set_linewidth(1.5) fig.text(0.57, 0.38, "doany-ret-v3 vs. Published", fontsize=16, fontweight="bold", color=GOLD, fontfamily="sans-serif") fig.text(0.57, 0.32, "Outperforms all baselines on our\ninternal eval suite by 5.8pp+", fontsize=14, color=WHITE, fontfamily="sans-serif", linespacing=1.4) fig.text(0.57, 0.22, "Competitive with MTEB/BEIR top models\nat significantly lower latency", fontsize=13, color=MUTED, fontfamily="sans-serif", linespacing=1.4) add_footer(fig, 6) return fig # ══════════════════════════════════════════════════════════════════════ # SLIDE 7 — Impact & Next Steps # ══════════════════════════════════════════════════════════════════════ def slide_impact(): fig = new_slide() fig.text(0.06, 0.88, "Impact & Next Steps", fontsize=36, fontweight="bold", color=WHITE, fontfamily="sans-serif") ax_line = fig.add_axes([0.06, 0.855, 0.24, 0.003]) ax_line.set_facecolor(ACCENT); ax_line.set_xticks([]); ax_line.set_yticks([]) for s in ax_line.spines.values(): s.set_visible(False) # Three action cards cards = [ ("1", "Update Production Benchmarks", "Replace all references to previous\nnumbers with corrected results", "Immediate"), ("2", "Rerun Q1 Ablation Study", "Repeat the ablation with the fixed\nindexing pipeline for accurate deltas", "This sprint"), ("3", "Publish Corrected Results", "Consider updating external-facing\nbenchmark documentation", "Next review"), ] x = 0.06 for num, title, desc, timeline in cards: # Card background ax_card = fig.add_axes([x, 0.22, 0.27, 0.55]) ax_card.set_facecolor(BG_CARD) ax_card.set_xticks([]); ax_card.set_yticks([]) for s in ax_card.spines.values(): s.set_color(BG_CARD); s.set_linewidth(0) # Number circle accent fig.text(x + 0.025, 0.71, num, fontsize=32, fontweight="bold", color=ACCENT, fontfamily="sans-serif") fig.text(x + 0.025, 0.62, title, fontsize=18, fontweight="bold", color=WHITE, fontfamily="sans-serif") fig.text(x + 0.025, 0.48, desc, fontsize=14, color=MUTED, fontfamily="sans-serif", linespacing=1.5) # Timeline badge fig.text(x + 0.025, 0.32, timeline, fontsize=13, fontweight="bold", color=GOLD, fontfamily="sans-serif", bbox=dict(boxstyle="round,pad=0.3", facecolor=BG, edgecolor=GOLD, linewidth=1)) x += 0.31 add_footer(fig, 7) return fig # ══════════════════════════════════════════════════════════════════════ # SLIDE 8 — Q&A # ══════════════════════════════════════════════════════════════════════ def slide_qa(): fig = new_slide() fig.text(0.5, 0.58, "Questions?", fontsize=56, fontweight="bold", color=WHITE, fontfamily="sans-serif", ha="center", va="center") fig.text(0.5, 0.44, "doany Retrieval Team | April 2026", fontsize=20, color=MUTED, fontfamily="sans-serif", ha="center") # Accent lines ax_top = fig.add_axes([0.35, 0.68, 0.30, 0.003]) ax_top.set_facecolor(ACCENT); ax_top.set_xticks([]); ax_top.set_yticks([]) for s in ax_top.spines.values(): s.set_visible(False) ax_bot = fig.add_axes([0.40, 0.38, 0.20, 0.003]) ax_bot.set_facecolor(ACCENT); ax_bot.set_xticks([]); ax_bot.set_yticks([]) for s in ax_bot.spines.values(): s.set_visible(False) add_footer(fig, 8) return fig # ══════════════════════════════════════════════════════════════════════ # BUILD PDF # ══════════════════════════════════════════════════════════════════════ def main(): builders = [ slide_title, slide_why, slide_headline, slide_recall, slide_latency, slide_related, slide_impact, slide_qa, ] print(f"Building {len(builders)}-slide PDF → {OUT_PDF.name}") with PdfPages(str(OUT_PDF)) as pdf: for i, build in enumerate(builders, 1): print(f" Slide {i}/{len(builders)}: {build.__name__}") fig = build() pdf.savefig(fig, facecolor=fig.get_facecolor(), dpi=200) plt.close(fig) size_mb = OUT_PDF.stat().st_size / (1024 * 1024) print(f"\nDone! {OUT_PDF} ({size_mb:.1f} MB)") if __name__ == "__main__": main()