{
  "docId": "019dd923-5e88-73ef-bd5d-0856e1444fb9",
  "docSlug": "8757f1b44ef7f176",
  "documentTitle": "2024 Air Street Capital The State of AI Report 2024",
  "authorId": "AirStreetCapital",
  "authorName": "Air Street Capital",
  "documentKindSlug": "consulting-deck",
  "documentKindLabel": "Consulting deck",
  "sourceTypeSlug": "vc_research",
  "sourceTypeLabel": "VC research",
  "presentationDate": null,
  "orientation": "landscape",
  "aspectRatio": 1.777,
  "pageNumber": 34,
  "pageCount": 213,
  "prevPage": 33,
  "nextPage": 35,
  "slideType": "industry_trends",
  "function": "present_solution",
  "density": "dense",
  "nDataPoints": 3,
  "notes": "The slide discusses a specific technical improvement in Retrieval-Augmented Generation (RAG) systems.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "flowchart"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0856e1444fb9/34",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0856e1444fb9#slide-34",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "They found that this approach leads to a reduction of top-20 retrieval failure rate of 35% (5.7% → 3.7%).",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d10-bee66e6c8b1b",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.35,
        "w": 0.38,
        "x": 0.58,
        "y": 0.41
      },
      "kind": "image",
      "text": "Contextual Retrieval Preprocessing diagram",
      "attrs": null,
      "subkind": "infographic",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "26a129e3-2396-4b6a-bb80-6389c7a674be",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.7,
        "w": 0.55,
        "x": 0.02,
        "y": 0.22
      },
      "kind": "list",
      "text": "Traditional RAG solutions usually involve creating text snippets 256 tokens at a time with sliding windows (128 overlapping the prior chunk). This makes retrieval more efficient, but significantly less accurate.\nAnthropic solved this using 'contextual embeddings', where a prompt instructs the model to generate text explaining the context of each chunk in the document.\nThey found that this approach leads to a reduction of top-20 retrieval failure rate of 35% (5.7% -> 3.7%).\nIt can then be scaled using Anthropic's prompt caching.\nAs Fernando Diaz of CMU observed in a recent thread, this is a great example of techniques pioneered on one area of AI research (e.g. early speech retrieval and document expansion work) being applied to another. Another version of “what is new, is old”.\nResearch from Chroma shows that the choice of chunking strategy can affect retrieval performance by up to 9% in recall.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "2076235c-2803-4191-bbe8-70af6ea58815",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "retrieval failure rate: 35%",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d10-c48fefd19106",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "quote",
      "text": "As Fernando Diaz of CMU observed in a recent thread, this is a great example of techniques pioneered on one area of AI research (e.g. early speech retrieval and document expansion work) being applied to another.",
      "attrs": null,
      "subkind": null,
      "toolName": "Authority citation",
      "toolSlug": "authority-citation",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d10-c24a8a3067a0",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.5,
        "x": 0.02,
        "y": 0.14
      },
      "kind": "title",
      "text": "Context proves a crucial driver of performance",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "f010d700-e854-4c05-8b0a-a2fe022dd1c3",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Action Titles",
      "slug": "action-titles",
      "agent": "Architect",
      "layer": "slide",
      "matchId": "019dd95a-1055-74e0-a958-b8e9f5efabc4",
      "evidence": "Title 'Context proves a crucial driver of performance'",
      "confidence": 88
    }
  ],
  "frameworks": [],
  "arcBeats": [
    {
      "to": 153,
      "from": 9,
      "beatId": "019dd95a-0682-776c-8e35-5f2398b8d1d0",
      "arcName": "The Triple Take",
      "arcSlug": "triple-take",
      "beatName": "The Facts (What)",
      "beatSlug": "triple-take-the-facts-what",
      "evidence": "Exec summary + Research + Industry sections inventory the year",
      "position": 1,
      "confidence": 78,
      "parentBeatName": "Setup",
      "parentBeatSlug": "setup"
    },
    {
      "to": 85,
      "from": 12,
      "beatId": "019dd95a-0682-776c-8e35-6d9d8ee06d0a",
      "arcName": "Voyage and Return",
      "arcSlug": "voyage-return",
      "beatName": "The Unknown",
      "beatSlug": "voyage-return-the-unknown",
      "evidence": "Research section explores frontier model uncharted territory",
      "position": 2,
      "confidence": 55,
      "parentBeatName": "Complication",
      "parentBeatSlug": "complication"
    }
  ],
  "loops": [
    {
      "to": 37,
      "from": 30,
      "name": "Pattern Hunter",
      "slug": "02-pattern-hunter",
      "bestFor": "Time-pressed audiences, building consensus, when data is strong",
      "matchId": "019dd95a-07fe-70ce-8d3e-2e9bfb3e3bea",
      "evidence": "Synthetic data, FineWeb, embeddings, RAG, power grid, JEST data curation",
      "position": 4,
      "objective": "Survey data, retrieval and training-infra advances",
      "structure": "Evidence A -> Evidence B -> Evidence C -> Pattern/Conclusion",
      "confidence": 72,
      "description": "Group multiple pieces of evidence that together point to a pattern or conclusion"
    }
  ],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}