How should documents become retrievable units without losing meaning or authority?

Versioned evidence corpus · deterministic retrieval fixtures

Lighthouse Policy Evidence Assistant. A fictional support assistant answers questions from a versioned product-policy corpus. It must retrieve current, authorized evidence, cite exact source spans, and abstain when evidence is missing or conflicting.

01 · Smallest useful mechanism

Preserve document structure, provenance, and lifecycle through ingestion.

Parsing and chunking decide which facts coexist, which headings survive, and whether a retrieved span can be traced back to an authoritative source. Stable identities and deletion propagation matter as much as token size.

A chunk is a derived evidence unit, not an anonymous bag of tokens.

02 · Experiment

Test the prediction

Deterministic retrieval fixture

This workbench uses inspectable ranking and evidence fixtures and injected failures. It does not make live calls or execute external effects.

Retrieval workbench · lighthouse-chunks-v2

Choose the chunk build that keeps a policy rule and its exception auditable.

Injected failure

The exception begins 12 tokens after a fixed 800-token boundary; the previous version must also be deleted from search.

Choose the implementation

Anonymous overlapping windowsUse 800 tokens with 400 overlap and store only the text and vector.One chunk per documentPreserve all relationships by embedding the complete policy.Structured child with parent contextKeep rule and exception together, inherit metadata, and retain exact spans.

The trace has not run.

Commit to an implementation, then inject the failure and inspect the actual state transitions.

03 · Implementation brief

Build it, break it, prove the outcome.

Implement: Define a chunk record, build manifest, stable ID rule, parent-child strategy, and deletion propagation path.
Break it on purpose: Split a rule from its exception, add repeated headers, change one ACL, replace the source, and delete the previous version.
Completion evidence: Golden fixtures preserve the clause, every chunk reconciles to one source span, access changes propagate, and no deleted chunk remains searchable.

Your learning artifact

Write the argument you would defend

Stored only in this browser. No account required; course reset does not delete it.

Your response

04 · Check your understanding

Which property makes a retrieved chunk usable as audit evidence?

Primary sources · verified 2026-08-02

What the fixture does not prove

Naive fixed windows can separate exceptions from rules, duplicate evidence through overlap, leak headers across tenants, and leave deleted content searchable after the source changed.

Common mistake: There is one universally optimal chunk size. Useful granularity depends on document structure, query type, embedding model, reranker, context budget, and the evidence span required by the product claim.

Transfer exercise: Split one policy by fixed tokens, headings, and parent-child sections; inspect which strategy keeps the exception attached to its rule.

OpenAI — Retrieval guideOfficial implementation contract for vector stores, semantic and hybrid search, attribute filters, rankers, thresholds, chunking, and result inspection.Anthropic — Contextual RetrievalVendor-authored experiments combining chunk context, embeddings, BM25, and reranking. The measured gains are useful hypotheses to reproduce on the target corpus.Lewis et al. — Retrieval-Augmented GenerationThe original RAG paper formalizing generation with parametric memory plus retrieved non-parametric evidence and explicit provenance goals.

Optional referenceDeep dive / reference chapterOpen the complete essay, diagrams, mathematics, exercises, glossary, and sources when you want more depth.on demand

Next: Once the corpus is trustworthy, candidate retrieval must combine lexical precision, semantic recall, filters, and query behavior without confusing scores.