Lab 2 — Local RAG

Fact-check status

Reviewed claim by claim on 2026-09-02 against the primary or official sources identified inline as S1 and S2. Recommendations are design guidance unless a cited standard makes them normative.

Key jargon

Term	Plain-language meaning
Corpus	The collection of source documents available for retrieval. S1
Index	A data structure that makes candidate retrieval efficient. S1
Retrieval query	The representation of the user's information need sent to search. S1
Ground-truth case	A question paired with known supporting evidence and expected answer properties. S1

Key concepts

Build a small measurable pipeline before adding frameworks or agent loops. S1
Inspect retrieved chunks for every test case; final-answer quality alone cannot diagnose retrieval failure. S1

Concept map

flowchart LR
    A["Prepare provenance-rich corpus"] --> B["Chunk embed and index"]
    B --> C["Retrieve and answer"]
    C --> D["Measure retrieval and grounding"]

Diagram semantics verified against S1 and S2 on 2026-09-02.

Goal

Answer questions from five synthetic Markdown documents using a local embedding index and model, with source IDs and abstention.

Build stages

Create documents with known facts, dates, conflicts, and access labels.
Preserve originals and generate a versioned ingestion manifest.
Compare two chunking strategies.
Build lexical and vector retrieval baselines.
Add context construction and generated answer.
Validate citations against exact source spans.
Test answerable, unanswerable, conflicting, and unauthorized cases.

Pass criteria

Required evidence appears in top-k retrieval for the agreed threshold.
Unauthorized document never reaches context.
Unsupported questions abstain.
Every answer identifies source document and span.

Continue

Use the preserved RAG-CAS notes as an implementation comparison, not as the evaluation definition. S1