Elastic
This account is verified because it's an official organization on X. Learn more@elastic
Where developers learn, build, and share. Your source for hands-on demos, cheat sheets, explainers and more.
186 Following 66K Followers 11K Posts
Latest posts 5
Your API's average latency is 80ms. Looks fine.
Now look closer: 98 requests took 20ms. 2 took 3 seconds.
Averages hide the tail. And the tail is real users, often your heaviest ones, because more requests means more chances to hit it.
Watch p95 and p99, not just the mean.
In Elastic APM, switch the latency chart from Average to 95th or 99th percentile and see who's actually waiting.
Now look closer: 98 requests took 20ms. 2 took 3 seconds.
Averages hide the tail. And the tail is real users, often your heaviest ones, because more requests means more chances to hit it.
Watch p95 and p99, not just the mean.
In Elastic APM, switch the latency chart from Average to 95th or 99th percentile and see who's actually waiting.
Ask semantic search for "error E1042" and you may get docs about errors in general.
Not the one with E1042.
Vectors capture meaning. Exact tokens like error codes, SKUs and names get blurred.
Keyword search (BM25) nails exact matches but misses meaning.
So run both, and merge the two ranked lists with Reciprocal Rank Fusion (RRF). Docs that rank high in either list rise to the top, with no need to normalize two very different score scales.
That's hybrid search.
Not the one with E1042.
Vectors capture meaning. Exact tokens like error codes, SKUs and names get blurred.
Keyword search (BM25) nails exact matches but misses meaning.
So run both, and merge the two ranked lists with Reciprocal Rank Fusion (RRF). Docs that rank high in either list rise to the top, with no need to normalize two very different score scales.
That's hybrid search.
How does search find one word in a billion documents in milliseconds?
It doesn't read the documents.
At index time, Elasticsearch builds an inverted index: every word mapped to the list of documents that contain it. Like the index at the back of a book.
Search "timeout" > jump to that entry > get the doc IDs > fed into BM25 for scoring.
The work moves from search time to write time. That trade is why search is fast.
It doesn't read the documents.
At index time, Elasticsearch builds an inverted index: every word mapped to the list of documents that contain it. Like the index at the back of a book.
Search "timeout" > jump to that entry > get the doc IDs > fed into BM25 for scoring.
The work moves from search time to write time. That trade is why search is fast.
Meta's Muse runs a 37.8M-param reranker on 2 CPU cores.
@rohanadwankar tore down the Muse agent VM. In its memory stack: jina-reranker-v1-turbo-en, 153 MB on disk.
Muse searches its users' memories in Postgres with vectors plus keywords. The Jina reranker orders those hits before the agent reads them.
Per the teardown, retrieval runs inside the VM, so memory chunks stay on the encrypted volume.
It's our English-only reranker from 2024. It still earns its place on a box with no GPU.
You can also run our latest @JinaAI_ models with Jina On-Prem. For self-contained and air-gapped deploys, with no outbound calls. .
Rohan's full teardown: go.es.io/4sxLHV7
@rohanadwankar tore down the Muse agent VM. In its memory stack: jina-reranker-v1-turbo-en, 153 MB on disk.
Muse searches its users' memories in Postgres with vectors plus keywords. The Jina reranker orders those hits before the agent reads them.
Per the teardown, retrieval runs inside the VM, so memory chunks stay on the encrypted volume.
It's our English-only reranker from 2024. It still earns its place on a box with no GPU.
You can also run our latest @JinaAI_ models with Jina On-Prem. For self-contained and air-gapped deploys, with no outbound calls. .
Rohan's full teardown: go.es.io/4sxLHV7
