{
  "schema_version": "1.1",
  "id": "s8:https://www.youtube.com/watch?v=r9OwPx_HoV0",
  "slug": "stop-chunking-like-it-s-2022-yuval-belfer-ai21-labs-1bsztja",
  "url": "https://feed7.dev/p/stop-chunking-like-it-s-2022-yuval-belfer-ai21-labs-1bsztja",
  "title": "Stop Chunking Like It's 2022 — Yuval Belfer, AI21 Labs",
  "why_included": "Fixed chunk sizes leave query-dependent retrieval quality on the table; multiscale indexes plus reciprocal rank fusion improved recall across four tested datasets at a memory cost.",
  "summary": "AI21 tested several chunk sizes on QMsum, NarrativeQA, Seinfeld, and FinanceBench. No fixed size consistently dominated; an oracle choosing per query showed a **20–40%** gap, while multiscale retrieval produced reported gains of roughly **10–40%** depending on the dataset.",
  "practical_implication": "Instead of committing to one window, index several sizes, query every index in parallel, map chunks back to document rankings, and merge them with **reciprocal rank fusion**. Treat chunking as a retrieval configuration that should be evaluated against real query shapes.",
  "agent_context": "AI21 tested several chunk sizes on QMsum, NarrativeQA, Seinfeld, and FinanceBench. No fixed size consistently dominated; an oracle choosing per query showed a **20–40%** gap, while multiscale retrieval produced reported gains of roughly **10–40%** depending on the dataset.\n\nInstead of committing to one window, index several sizes, query every index in parallel, map chunks back to document rankings, and merge them with **reciprocal rank fusion**. Treat chunking as a retrieval configuration that should be evaluated against real query shapes.\n\nThe method duplicates indexes and was reported to require roughly **2–5x memory**. The tested sizes were chosen somewhat arbitrarily, and the talk leaves the best number of scales and ranking fusion method open.",
  "source": {
    "name": "AI Engineer",
    "url": "https://www.youtube.com/watch?v=r9OwPx_HoV0",
    "published_at": "2026-09-16T17:00:06.000Z"
  },
  "source_class": "video",
  "content_type": "Video",
  "layer": "context",
  "domains": [
    "research",
    "data"
  ],
  "topics": [
    "retrieval",
    "context-engineering"
  ],
  "verification": {
    "status": "source_linked",
    "label": "Source Linked",
    "method": "source_feed",
    "verified_at": null
  },
  "uncertainty": [
    "The method duplicates indexes and was reported to require roughly **2–5x memory**. The tested sizes were chosen somewhat arbitrarily, and the talk leaves the best number of scales and ranking fusion method open."
  ],
  "connected_context": {
    "meaning": "Multiscale retrieval turns chunk size from a fixed ingestion choice into a query-dependent ranking problem. It strengthens the prior case for workload-specific retrieval evaluation by showing that no tested size consistently wins, while supplying a practical hedge through parallel indexes and rank fusion. That hedge has a material storage cost, and the unresolved number of scales and fusion method mean it should be tested against simpler retrieval and full-context baselines.",
    "corpus_size": 812,
    "generated_at": "2026-09-19T09:05:18.698Z",
    "connections": [
      {
        "title": "Teaching Nemotron Greek: Mining a Corpus, Adapting Retrieval, and Grounding Generation for Modern Greek across Specialist Domains",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2608.05138v1",
        "feed7_url": "https://feed7.dev/p/2608-05138v1-0bvu6le",
        "reason": "The Greek RAG study independently shows that retrieval choices vary by language and domain, reinforcing the need to evaluate multiscale chunking on actual corpora and queries rather than assume a universal configuration."
      },
      {
        "title": "infiniflow/ragflow",
        "source_name": "GitHub",
        "source_url": "https://github.com/infiniflow/ragflow",
        "feed7_url": "https://feed7.dev/p/ragflow-0b0v2o1",
        "reason": "RAGFlow provides an implementation surface where multiple chunking and ranking configurations could be tested, but its packaged pipeline does not remove the need to justify multiscale indexing’s added storage and complexity."
      },
      {
        "title": "Context Engineering in 2026 — Louis-François Bouchard, Omar Solano & Samridhi Vaid, Towards AI",
        "source_name": "AI Engineer",
        "source_url": "https://www.youtube.com/watch?v=WP3hjUXd918",
        "feed7_url": "https://feed7.dev/p/context-engineering-in-2026-louis-francois-bouchard-omar-solano-samridhi-1dnlyr0",
        "reason": "The tutor evaluation supplies an important competing baseline: when full cached history is cheaper and preserves more detail, multiscale retrieval must demonstrate value against retaining the complete context."
      },
      {
        "title": "ReContext: Recursive Evidence Replay as LLM Harness for Long-Context Reasoning",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2607.02509v1",
        "feed7_url": "https://feed7.dev/p/2607-02509v1-11vodps",
        "reason": "ReContext offers a different training-free response to long inputs by replaying query-relevant evidence, contrasting with multiscale retrieval’s decision to improve evidence selection at indexing and ranking time."
      }
    ]
  },
  "lifecycle": "Current",
  "published_at": "2026-09-16T17:00:06.000Z",
  "modified_at": "2026-09-16T17:00:06.000Z",
  "supersedes": [],
  "expires_at": null,
  "formats": {
    "html": "https://feed7.dev/p/stop-chunking-like-it-s-2022-yuval-belfer-ai21-labs-1bsztja",
    "json": "https://feed7.dev/p/stop-chunking-like-it-s-2022-yuval-belfer-ai21-labs-1bsztja.json",
    "markdown": "https://feed7.dev/p/stop-chunking-like-it-s-2022-yuval-belfer-ai21-labs-1bsztja.md"
  }
}