← All problem sets

SCADS 2026 / Problem 06

Agentic retrieval from selected resources

Build context from a transcript.

Research planning · constrained retrievalNot started

The task

How could an agent enrich a presentation transcript with credible context about its speakers, topics, venue, and related talks from a predefined set of sources?

Projected solution

I propose an allowlisted research agent that plans its next lookup from the evidence it finds, resolves ambiguous identities, and produces a source-linked contextual dossier.

The proposed workflow starts with transcript anchors and an explicit source registry. A bounded planner chooses lookups, tracks unresolved questions, and cross-checks claims. A fixed-workflow baseline and held-out transcripts would test whether adaptive planning improves useful coverage within the same budget.

Proposed workflow

Proposed: source registry · agent planner · evidence ledger
  1. 01

    Anchor

    Identify transcript claims, people, and topics with precise transcript locations.

  2. 02

    Investigate

    Choose permitted resources dynamically and record the basis for each lookup.

  3. 03

    Verify

    Cross-reference claims and assemble a dossier with source and transcript citations.

Implementation roadmap

Proposed milestones / Not started
  1. Define and freeze. Specify the task, data rights, baselines, evaluation criteria, and reserved test material.
  2. Build the smallest end-to-end study. Implement the proposed workflow with traceable inputs and reproducible outputs.
  3. Compare and stress-test. Run controlled baseline comparisons and ablations; inspect failure modes and resource use.
  4. Reproduce and report. Repeat the study from a clean environment and publish measured findings with supporting artifacts.

The research proposal expands these milestones into project-specific methods, experimental controls, and deliverables.

Methods / evidence / reproduction

Read the research.

The proposed evaluation compares adaptive retrieval with fixed retrieval under matched request and time budgets. Coverage, citation support, identity errors, and reproducibility would be measured on held-out presentations.

The proposal develops the research question, proposed method, evaluation design, and phased roadmap. Its GitHub directory contains the planning documents. No implementation or experimental results are claimed.