From ecbc0a7fea3f7c5bcd822f94a64a99bda733b1b1 Mon Sep 17 00:00:00 2001 From: robertturner <143536791+robertdhayanturner@users.noreply.github.com> Date: Mon, 26 Aug 2024 20:04:12 +0200 Subject: [PATCH] Revert "Update rag-application-communication-system.md" --- docs/articles/rag-application-communication-system.md | 11 +++++------ 1 file changed, 5 insertions(+), 6 deletions(-) diff --git a/docs/articles/rag-application-communication-system.md b/docs/articles/rag-application-communication-system.md index 84ce8baa5..7de6a2fbb 100644 --- a/docs/articles/rag-application-communication-system.md +++ b/docs/articles/rag-application-communication-system.md @@ -96,10 +96,10 @@ In practical terms: In this time of LLMs, older information retrieval methods and indicators continue to hold a lot of unrealized value, especially now that it's possible to generate/extract many key data features at scale. Jo Kristian Bergum from Vespa, for example, has [convincingly demonstrated](https://blog.vespa.ai/improving-retrieval-with-llm-as-a-judge/) how classic info retrieval evaluation design and metrics (precision at k, recall) can be effectively repurposed using emerging practices in AI, such as LLM-as-a-Judge - grounded on a small but scalable relevant dataset. Intensive data work that would have been available only to large scale organizations is now scalable with far fewer resources. >   ->**GOING HYBRID**             ->    - *Indexation*: traditional keyword matching + modern embedding-based similarity ->    - *Searching*: keyword-based search + vector search ->    - *Evaluation*: precision at k, recall + LLM-as-a-judge +>**GOING HYBRID** +>- *Indexation*: traditional keyword matching + modern embedding-based similarity +>- *Searching*: keyword-based search + vector search +>- *Evaluation*: precision at k, recall + LLM-as-a-judge >   Generative AI within a RAG communication system shouldn't be looking to replace the classic approaches of retrieval evaluation; it should instead reshape their logistics to take full advantage of them. @@ -112,8 +112,7 @@ A proper RAG communication system should treat data no longer as a passive refer 2. constantly circulated across different flows >   -> A good RAG comm system includes:             ->     bad data + classifiers + synthetic data curation +> A good RAG comm system includes: bad data + classifiers + synthetic data curation >   ### 3.1 You need bad data