๐ถ๏ธ Hot take: I'm tired of RAG systems that just grab the first...
Victoria Slocum@victorialslocum
14 views
Aug 26, 2025
~2 min read
Advertisement
1
๐ถ๏ธ Hot take: I'm tired of RAG systems that just grab the first document and call it done.
๐๐ด๐ฒ๐ป๐๐ถ๐ฐ ๐ฅ๐๐ is probably the only way RAG is ๐ข๐ค๐ต๐ถ๐ข๐ญ๐ญ๐บ going to be useful.
But whatโs actually the difference?
๐ง๐ต๐ฒ ๐ฃ๐ฟ๐ผ๐ฏ๐น๐ฒ๐บ ๐๐ถ๐๐ต ๐ก๐ฎ๐ถ๐๐ฒ ๐ฅ๐๐:
Traditional RAG pipelines are essentially one-shot systems, they retrieve context once and hope for the best. They're limited to a single external knowledge source and have zero reasoning about retrieval quality or the data environment theyโre working with.
The difference in ๐๐ด๐ฒ๐ป๐๐ถ๐ฐ ๐ฅ๐๐ systems comes down to three capabilities:
๐ญ. ๐ ๐ฒ๐บ๐ผ๐ฟ๐: Unlike naive RAG that forgets everything between queries, agentic systems maintain both short-term and long-term memory. This means they can learn from previous interactions and build context over time.
๐ฎ. ๐ง๐ผ๐ผ๐น๐: While naive RAG is stuck with one knowledge source, agentic RAG can access multiple external tools like vector dbs, web searches, APIs, calculators, whatever the task requires, and also dynamically formulate *how* to query that tool. This makes it a lot more robust in dealing with actual human usage.
๐ฏ. ๐ฆ๐ฒ๐น๐ณ-๐๐ฒ๐ฎ๐น๐ถ๐ป๐ด ๐ฅ๐ฒ๐ฎ๐๐ผ๐ป๐ถ๐ป๐ด: Agentic systems can reflect on their retrieval quality, critique their own responses, and iterate until they get better results. They actually ๐กโ๐๐๐ about whether the retrieved context is good enough, whether theyโve responding to the userโs query to the best of their ability.
The process becomes iterative - the agent retrieves, reasons, validates, and refines until it completes the task properly, and with the right architecture implementation, will improve over time.
So yes sure there are still plenty of applications for Ask-Retrieve-Generate workflows, but if we actually want robust systems that provide value, agentic RAG is going to be the future.
After building our naive RAG project Verba, we realized pretty quickly it had some major limitations. So thatโs exactly why we built our new agentic RAG framework, Elysia, thatโs an entirely open source, pip-installable python package.
Read more about how we built a full end-to-end agentic system: weaviate.io/blog/elysia-agโฆ
๐๐ด๐ฒ๐ป๐๐ถ๐ฐ ๐ฅ๐๐ is probably the only way RAG is ๐ข๐ค๐ต๐ถ๐ข๐ญ๐ญ๐บ going to be useful.
But whatโs actually the difference?
๐ง๐ต๐ฒ ๐ฃ๐ฟ๐ผ๐ฏ๐น๐ฒ๐บ ๐๐ถ๐๐ต ๐ก๐ฎ๐ถ๐๐ฒ ๐ฅ๐๐:
Traditional RAG pipelines are essentially one-shot systems, they retrieve context once and hope for the best. They're limited to a single external knowledge source and have zero reasoning about retrieval quality or the data environment theyโre working with.
The difference in ๐๐ด๐ฒ๐ป๐๐ถ๐ฐ ๐ฅ๐๐ systems comes down to three capabilities:
๐ญ. ๐ ๐ฒ๐บ๐ผ๐ฟ๐: Unlike naive RAG that forgets everything between queries, agentic systems maintain both short-term and long-term memory. This means they can learn from previous interactions and build context over time.
๐ฎ. ๐ง๐ผ๐ผ๐น๐: While naive RAG is stuck with one knowledge source, agentic RAG can access multiple external tools like vector dbs, web searches, APIs, calculators, whatever the task requires, and also dynamically formulate *how* to query that tool. This makes it a lot more robust in dealing with actual human usage.
๐ฏ. ๐ฆ๐ฒ๐น๐ณ-๐๐ฒ๐ฎ๐น๐ถ๐ป๐ด ๐ฅ๐ฒ๐ฎ๐๐ผ๐ป๐ถ๐ป๐ด: Agentic systems can reflect on their retrieval quality, critique their own responses, and iterate until they get better results. They actually ๐กโ๐๐๐ about whether the retrieved context is good enough, whether theyโve responding to the userโs query to the best of their ability.
The process becomes iterative - the agent retrieves, reasons, validates, and refines until it completes the task properly, and with the right architecture implementation, will improve over time.
So yes sure there are still plenty of applications for Ask-Retrieve-Generate workflows, but if we actually want robust systems that provide value, agentic RAG is going to be the future.
After building our naive RAG project Verba, we realized pretty quickly it had some major limitations. So thatโs exactly why we built our new agentic RAG framework, Elysia, thatโs an entirely open source, pip-installable python package.
Read more about how we built a full end-to-end agentic system: weaviate.io/blog/elysia-agโฆ
