← sources

The Vector DB is an Overkill for almost all RAG applications and response is only as accurate as Relational database on average!

source: reddit (reddit:developersIndia)published 2026-11-07 03:56ingested 2026-11-08 02:36

retained text

Research paper: Is Grep All You Need? How Agent Harnesses Reshape Agentic Search

Main Result: In their experiments (evaluating on a sample from LongMemEval across custom and native CLI harnesses like Claude Code, Codex, and Gemini CLI), grep generally yielded higher accuracy than vector retrieval—especially when facts are literal, anchored in names, dates, or code strings.

[https://arxiv.org/pdf/2605.15184](https://arxiv.org/pdf/2605.15184)

Comparison table provided by Chatgpt. (Illustrative relative cost model, not measured industry averages. Assumes comparable workloads, LLM models and deployment scale. Actual ratios depend heavily on implementation.)

|Architecture|All Queries Combined accuracy |Build Cost|Running Cost|Response Time|General Knowledge|How-to / Instructions|Recommendations / Decisions|Complex Research / Analysis|Personalised Queries|
|:-|:-|:-|:-|:-|:-|:-|:-|:-|:-|
|Traditional RAG (Vector DB)|75%|3x|3x|Fast|85%|80%|75%|65%|70%|
|Agentic RAG / Hybrid|89%|5x|5x|Slow|92%|90%|88%|85%|90%|
|LLM + Relational DB (SQL + FTS)|72%|1x|1x|Very Fast|95%|65%|60%|50%|90%|
|LLM + Grep (Agentic Loop)|74%|0x|2x|Moderate–Slow|90%|85%|70%|65%|60%|

What has been your experience and accuracy statistics?

document fetched 2026-11-08 00:22 · canonical https://reddit.com/r/developersIndia/comments/1xpwulp/the_vector_db_is_an_overkill_for_almost_all_rag