[HN Gopher] Rerank 3: A new foundation model for efficient enter...
___________________________________________________________________
Rerank 3: A new foundation model for efficient enterprise search
and retrieval
Author : bguberfain
Score : 35 points
Date : 2024-04-11 17:38 UTC (5 hours ago)
(HTM) web link (txt.cohere.com)
(TXT) w3m dump (txt.cohere.com)
| bigcat12345678 wrote:
| El5 me what is rerank model? Why 4k context window size is
| considered large?
| razodactyl wrote:
| Imagine you have 100 documents in a database and you "query"
| the documents and return 20 candidate results.
|
| Similarity gave you 20 results but Re-ranking sorted those
| results further providing relevance.
|
| That 4K is per document.
|
| Edit: With sorted relevance, you can drop the lower scoring
| documents according to the model's confidence that the
| information in the subset is adequate to answer the query.
| bigcat12345678 wrote:
| 20 results, each have 4k snippet, feed to this ranking model,
| ranking model produces a score based on query
|
| Is this a correct understanding?
| dvt wrote:
| Being as charitable as possible here, and Rerank 3 might be the
| bee's knees, but the examples are absolutely _awful_. Do you
| really need to use embeddings + a large language model to search
| for "action" and "Christian Bale" in two columns[1]?
|
| Your interface can literally just be two dropdowns. I'd like to
| see things like "the actor that played the Joker in that movie
| about Bob Dylan" if you're really trying to flex your semantic
| search muscles.
|
| [1]
| https://colab.research.google.com/drive/1sKEZY_7G9icbsVxkeEI...
| esafak wrote:
| Someone correct me if I'm mistaken, but Cohere appears to be
| using BM25 and semantic search (Embed Multilingual) individually
| as baselines in order to look better. A more suitable baseline
| would be the Reciprocal rank fusion (RRF) of BM25 and semantic
| search. And those latencies seem high; seconds to rerank?
___________________________________________________________________
(page generated 2024-04-11 23:02 UTC)