Cortex

Architecture

A query service embeds or tokenizes the question, retrieves candidate passages from a document store, optionally reranks them, then synthesizes an extractive answer strictly from selected evidence. An evaluation module scores groundedness and source coverage. This portfolio build uses a local lexical retriever over bundled sample documents—no model API is called.

Query → tokenize / expand → retrieve top-k passages → rerank by strategy → extractive synthesis → cite sources → evaluate overlap with supporting passages.

Decisions

• — Keep the answer extractive so every sentence can be traced to a passage.
• — Expose retrieval strategy and k as first-class controls rather than hidden hyperparameters.
• — Separate gold-supported answers for curated questions from fallback extractive synthesis for free-text queries.

Trade-offs

• — Lexical retrieval is inspectable and deterministic, but it will miss paraphrases that a dense embedder would catch.
• — Extractive answers are honest without an LLM, and they are less fluent than generative copy.
• — A tiny corpus makes ranking differences obvious; it is not a substitute for production-scale evaluation.

Failure modes

• — Queries outside the corpus return low coverage and a refused or weakly supported answer.
• — Raising k can dilute precision if later passages are only loosely related.
• — A title-heavy rerank can overfit to document names instead of passage content.

Evaluation

Groundedness estimates how much of the answer is attested in retrieved text. Source coverage estimates how many of the cited documents actually appeared in the retrieved set. Both are illustrative checks on demo data, not published quality benchmarks.

Production extensions

Proposed — not implemented in this portfolio.

• — Swap the lexical index for embeddings and a vector store.
• — Add a real LLM with citation constraints and a hallucination grader.
• — Persist query traces for offline evaluation sets.