Skip to content

Section 1 notes: the traversal problem

Open — needs literature research.

Key claims to support

  1. Current AI goal-resolution systems use scan as their primary navigation mode (RAG, web search agents, long-context models).
  2. The cost of scan scales with information-space size, not goal complexity.
  3. For goals with determinate answers in structured knowledge, this is wasteful by construction — the waste is load-bearing, not incidental.

What to research

  • Published token/cost measurements for RAG pipelines on structured-domain queries
  • Papers comparing long-context vs. structured-retrieval cost on matched tasks
  • Empirical work on how many documents RAG systems evaluate before converging on an answer ("retrieval depth" studies if such exist)
  • Cost breakdowns for frontier model API usage by goal type

Candidate sources

  • RAGAS, ARES, RECALL benchmarks — may have per-query retrieval statistics
  • LlamaIndex, LangChain documentation and benchmark reports
  • Papers on "retrieval-augmented generation" efficiency (search: "RAG efficiency" "RAG cost" "token consumption retrieval")
  • GPT-4 / Claude cost calculator papers / analyses from practitioners

Framing note

The argument should not attack RAG as a technology — it is well-suited for open-domain synthesis tasks. The argument is narrower: for goals that are anchorable to structured knowledge, RAG's scan mode is a mismatch, and the mismatch has measurable cost. The paper should be explicit that this is a claim about a subset of goals, not all goals.

Open questions

  • Is there published work specifically comparing structured graph lookup vs. embedding retrieval on knowledge graph question answering (KGQA)? This could be the closest existing empirical comparison.
  • "Retrieval cost" in the literature often means time or FLOPS, not tokens. How do we bridge to token cost as the unit of comparison?