July 26, 2026
Week 1 Updates: v2 ready
Week 1: Running on a new index version, Books, Videos, Blog and more search updates.
Purili v2 Is Now Live
Purili has fully moved to its v2 search system.
The old runtime fallback has been removed, the production rollout is complete, and the public search stack now runs entirely on schema v2. The previous index and binaries are retained offline for recovery, but they are no longer part of normal search traffic.
Better ranking and retrieval
One of the largest improvements is passage-aware reranking.
Instead of judging pages mainly by generic summaries, Purili selects the text passage most relevant to the query before applying its quality and ranking signals. This gives the ranking system better context and helps surface pages that actually discuss the requested subject.
Purili also supplements broad retrieval with bounded constraint searches for longer queries. Technical searches can recognize index-backed compound terms, such as onnx runtime and onnxruntime, while low-coverage matches are demoted. Known navigational searches lead directly to verified official destinations, even when a JavaScript-heavy homepage is difficult to crawl.
Stronger search results
The latest comparable 500-query evaluation shows the progress since the earlier v2 build:
| Metric | Previous v2 | Current v2 |
|---|---|---|
| Strict judged score | 50.9 / 100 | 100.0 / 100 |
| Retrieval-health diagnostic | 73.0 / 100 | 75.2 / 100 |
| Top-10 topical results | 46.9% | 51.6% |
| Fully topical top-10 pages | 24.2% | 30.0% |
| Zero-result queries | 0 | 0 |
The 100% strict score applies to the 35 queries with explicit expected-result judgments. Across the complete 500-query inventory, the broader automated retrieval-health diagnostic reached 75.2.
The final 131-query production validation retained 55.7% top-1, 60.3% top-3, and 77.9% good@3. With the main crawler active, latency measured 425 ms at p50 and 2,098 ms maximum.
Automated evaluation remains only a proxy. For example, it does not consider the joined token onnxruntime equivalent to the separate words onnx runtime, even when the returned documentation is clearly relevant. Purili therefore uses explicit judgments and manual result inspection alongside automated scoring.
Smarter result pages
Purili now provides more useful information directly on the results page without relying on generative AI.
Instant answers use typed query understanding, local Wikipedia knowledge, structured data, trusted indexed excerpts, and a high-confidence built-in knowledge layer. The current production instant-answer suite passes 69 out of 69 checks.
Exact infocards remain visible reliably. Relevant books, videos, news, official destinations, FAQ answers, local details, and interactive tools can appear alongside normal web results when they match the query intent. Duplicate regional services and repeated parameter variants are collapsed so they do not occupy the result page several times.
A much larger index
Purili currently searches more than:
- 245 million web and news documents
- 40 million books
- 70 million videos
Books and videos are available both within relevant web searches and through their dedicated search pages.
What comes next
Complex, comparison, and developer-oriented queries remain the largest areas for improvement. A large index is only useful when the retrieval system can consistently find the correct URLs inside it.
The next measurable goals are:
- Increase automated top-10 topicality beyond 60%
- Raise fully topical result pages above 40%
- Expand query decomposition and authoritative documentation retrieval
- Build pooled human relevance judgments for more reliable NDCG@10 and MRR
- Continue improving result quality without sacrificing search latency
With v2 live, Purili has a simpler production architecture, a substantially larger index, stronger direct answers, and a better foundation for the next round of search-quality improvements.