# You're doing RAG wrong — Akshay > [!question] Lingering Questions > - Does "one question, one answer" atomization lose multi-hop or relational knowledge that a wider text chunk would have preserved? > - The 40x/3x/2.3x figures cited are single-vendor (Blockify) and unreproduced — worth independently benchmarking before treating them as anything more than a hypothesis. > [!link] Cross Reference > - [[3. Resources/Permanent Notes/Retrieval Quality Is Set by the Unit of Knowledge Not the Retrieval Algorithm]] — this source's central claim, already captured as a Permanent note with this thread as its origin. #xref > - [[3. Resources/Permanent Notes/Concentration Beats Coverage Under Retrieval]] — same underlying instinct: fewer, canonical units outperform broad redundant coverage at retrieval time. #xref > [!quote] Vocabulary > - **IdeaBlock** — the thread's proposed retrieval unit: one question, one validated answer, typed governance fields, on a single object — replacing the arbitrary text chunk. ## Bib capsule Akshay (@akshay_pachaar), "You're doing RAG wrong," X/Twitter thread, 2026. https://twitter.com/akshay_pachaar/status/2052743644411765230 ## Author's thesis (Adler Q1+Q2) Every RAG pipeline starts from an unexamined assumption — that a chunk of text is the right unit of knowledge to embed — when a chunk is really a structurally neutral container with no idea boundary, version, or access state. The thread argues retrieval quality is decided upstream of the retriever: embed atomic question-answer packets instead of prose windows, and deduplicate near-identical chunks down to canonical claims, and accuracy rises even as the corpus shrinks. This is disclosed vendor content (promoting a product called Blockify), so the specific benchmark numbers should be treated as a hypothesis to reproduce, not an established result. ## Key arguments - The text chunk is the wrong unit of knowledge to embed — it carries no idea boundary, version, or access-control state. - Embedding a question-answer packet makes retrieval structural rather than merely semantic, since user queries are already phrased as questions. - Redundant near-duplicate chunks compete for the same region of embedding space and dilute retrieval probability mass; deduplicating to canonical claims measurably shrank the corpus while raising vendor-reported accuracy. ## Claim atoms ### The text chunk is the wrong unit of knowledge to embed > [!quote] ID:rw_pending > "Every RAG pipeline starts with the same assumption: a chunk of text is the right unit of knowledge to embed. That assumption is almost never examined." **Where/why**: Opening thesis; reframes retrieval failure as an upstream data-modeling problem rather than a search-tuning one. Already promoted to a Permanent note. **Tags**: #xref #aphorisms ### Embed a question-answer packet, not a prose window > [!quote] ID:rw_pending > "When your index stores answers to questions, the match becomes structural, not just semantic." **Where/why**: The IdeaBlock — one question, one validated answer, typed governance fields, on one object. User queries are already questions, so question-to-answer matching beats hoping the right paragraph floats up. **Tags**: #xref #articleseed ### Deduplicating to canonical claims raises accuracy while shrinking the corpus > [!quote] ID:rw_pending > "fifteen near-duplicates of the same paragraph create fifteen competing vectors in the same region of embedding space. Retrieval distributes probability mass across all of them" **Where/why**: Vendor-reported: 2,042 → 1,200 blocks; 88,877 → 44,537 words; +13.55% vector accuracy (unsourced beyond the vendor — flag as a claim to verify). "Your vector index isn't a hard drive you want to fill. It's a retrieval surface, and redundancy degrades it." **Tags**: #xref ## Permanent candidates - **Retrieval quality is set by the unit of knowledge, not the retrieval algorithm** — already exists as [[3. Resources/Permanent Notes/Retrieval Quality Is Set by the Unit of Knowledge Not the Retrieval Algorithm]], created from this thread. Not duplicated here. - The Q&A-packet and deduplication atoms are both mechanisms *implementing* that same upstream claim rather than distinct reusable claims — not split into separate Permanent notes. ## Processing checklist - [x] All claim atoms have header + verbatim quote + where/why - [x] Triage callouts populated for every tagged highlight - [x] Checked existing Permanents first — found direct coverage, linked rather than duplicated - [x] Connected into MOC — [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] ## Links - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — retrieval-architecture decisions are load-bearing for any AI-assisted knowledge system - [[3. Resources/Permanent Notes/Retrieval Quality Is Set by the Unit of Knowledge Not the Retrieval Algorithm]] — because evidence-for; this thread is that claim's origin source - [[3. Resources/Readwise/Tweets/You're Doing RAG Wrong]] — source document