alete-ai/web-cache-similarity-benchmark

Web Cache Semantic Similarity Benchmark

This dataset was created by Alete to benchmark the effectiveness of modern semantic vector caching against traditional cryptographic string hashing (SHA-256) directly on the live web, incorporating both time-decay and browser-hydration variables.

Naïve string hashing fails on the modern web because minor changes in DOM structures—such as dynamic ad placements, changing comment widgets, timestamp updates, social link reorderings, or tracking query parameters—alter the byte signature completely. This causes false negatives (cache misses), forcing backend LLM extraction pipelines to consume extra tokens and reprocess pages that have not fundamentally changed.

Furthermore, this benchmark introduces a crucial engineering challenge: environmental layout hydration noise. When page HTML is retrieved via static HTTP clients (like requests.get) versus dynamic headless browsers (like Chrome via Firecrawl), JavaScript execution prepends/appends dynamic recommendation widgets and cookie consent frames inside the article container. This completely alters the DOM order and layout structure.

This benchmark evaluates 50 live articles across three snapshots ($T_1$, $T_2$, and $T_3$) representing different time intervals and crawling environments:

• $T_1$: Baseline direct HTTP fetch (requests.get at 11:44:08-04:00).
• $T_2$: Follow-up direct HTTP fetch (requests.get at 15:17:12-04:00 — 3.5 Hours later).
• $T_3$: Browser-hydrated Chrome fetch (Firecrawl API at 16:54:38-04:00 — 5.0 Hours later).

---

Dataset Schema

Each row contains the following metrics and content snapshots:

Field Name	Type	Description
url	string	The live publication URL of the news article.
source	string	The news publisher source (fox, bbc, guardian).
timestamp_t1	string	Timestamp for the baseline direct HTTP crawl ($T_1$).
timestamp_t2	string	Timestamp for the follow-up direct HTTP crawl ($T_2$).
timestamp_t3	string	Timestamp for the browser-hydrated Chrome crawl ($T_3$).
markdown_t1	string	Markdown compiled from raw HTML at $T_1$ using MDream minimal preset.
markdown_t2	string	Markdown compiled from raw HTML at $T_2$ using MDream minimal preset.
markdown_t3	string	Markdown compiled from hydrated Chrome HTML at $T_3$ using MDream.
digest_t1	string	M-Dream AST digest (Title + Headings + Lead 800 chars) for $T_1$.
digest_t2	string	M-Dream AST digest for $T_2$.
digest_t3	string	M-Dream AST digest for $T_3$.
tokens_digest	integer	Token footprint of the digest at $T_1$.
tokens_full	integer	Token footprint of the full markdown body at $T_1$.
sha_full_match	boolean	True if SHA256(markdown_t1) == SHA256(markdown_t2) (3.5-hour HTTP-to-HTTP match).
sha_digest_match	boolean	True if SHA256(digest_t1) == SHA256(digest_t2) (3.5-hour HTTP-to-HTTP digest match).
digest_similarity	float	Cosine similarity score of $T_1$ vs $T_2$ digests using Vertex AI text-embedding-005.
full_similarity	float	Cosine similarity score of $T_1$ vs $T_2$ full markdown using text-embedding-005.
sha_t1_t3_match	boolean	True if SHA256(markdown_t1) == SHA256(markdown_t3) (HTTP vs Browser match).
sha_t1_t3_digest_match	boolean	True if SHA256(digest_t1) == SHA256(digest_t3) (HTTP vs Browser digest match).
t1_t3_digest_similarity	float	Cosine similarity score of $T_1$ vs $T_3$ digests.
t1_t3_full_similarity	float	Cosine similarity score of $T_1$ vs $T_3$ full markdown.
sha_t2_t3_match	boolean	True if SHA256(markdown_t2) == SHA256(markdown_t3).
sha_t2_t3_digest_match	boolean	True if SHA256(digest_t2) == SHA256(digest_t3).
t2_t3_digest_similarity	float	Cosine similarity score of $T_2$ vs $T_3$ digests.
t2_t3_full_similarity	float	Cosine similarity score of $T_2$ vs $T_3$ full markdown.

---

3-Snapshot Cache Benchmark Results

The empirical results show the performance of caching under time decay and dynamic environment differences:

1. Pure Time-Decay Caching (T1 vs T2: 3.5 Hours, Direct HTTP to HTTP)

• SHA-256 Full Markdown Hits: 20.0% (80.0% cache misses due to dynamic layout adjustments).
• SHA-256 M-Dream Digest Hits: 92.0% (AST container isolation blocks most page noise).
• Digest Embedding Hits (threshold 0.92): 98.0% (Near-frictionless semantic cache hit rate).

2. Environmental Layout Hydration Caching (T1 vs T3: 5.0 Hours, HTTP vs Browser)

• SHA-256 Full Markdown Hits: 0.0% (100.0% misses. Dynamic javascript pre/post-pends recommended containers).
• SHA-256 M-Dream Digest Hits: 0.0% (100.0% misses. Consent banners and ad injection shift the DOM).
• Digest Embedding Hits (threshold 0.92): 26.0% (Vector distances expand; average similarity drops to 0.7965).
• Key Finding: Comparing a raw static HTTP scrape against a dynamic browser DOM breaks exact byte hashes 100% of the time, and expands vector space distances significantly. To achieve high vector hits (98%+), the pipeline must enforce Client-Backend Environmental Uniformity (using the same browser extension scraping agent for both baseline and lookup queries) or adjust thresholds dynamically to 0.78 when crossing environments.

---

How to Use the Dataset

You can load this dataset programmatically using the Hugging Face datasets library:

from datasets import load_dataset

dataset = load_dataset("alete-ai/web-cache-similarity-benchmark")
print(dataset["train"][0])