An engineer searches for 'wait after reset goes away,' while the specification says 'reset deassertion recovery interval.' The passage may be applicable even though the wording differs. Conversely, a search for a literal error code such as SYNTH_E17 should preserve that exact token; a generally related paragraph about synthesis is not an adequate replacement. Keyword and semantic retrieval address these cases differently. Learn their failure modes before treating either ranking as an answer.
Keyword retrieval works with words or tokens present in the text. Even a simple set-overlap implementation can be useful for a small collection with controlled terminology. More developed text search can handle phrases, field weights, and ranking. Semantic retrieval uses an embedding model to represent text for similarity comparison, which can help find paraphrases. Similarity does not mean the passage proves the requested fact. Neither approach automatically knows which project, revision, operating mode, or corner governs the question.
Question | Candidate passage | What to check |
|---|---|---|
reset release | Reset release requires three rising clock edges... | Exact shared terms make a useful keyword candidate. |
wait after reset goes away | Reset deassertion recovery interval... | A synonym expansion or semantic retriever might help; inspect the actual passage. |
SYNTH_E17 | General synthesis warning handling | Related subject matter does not establish the exact error-code meaning. |
read latency in 2.0 | Read latency in 1.0 | Reject the wrong revision before ranking eligible evidence. |
For a transparent baseline, lowercase the words, preserve identifier underscores, and count query terms found in each passage. If the required evidence terms are reset and release, both words must appear for the lab's candidate gate; the query words only rank passages that pass. This rule is easy to inspect but brittle: a valid passage using deassertion instead of release will be missed. Adding a reviewed synonym can help, but unrestricted substitutions may collapse distinctions such as asynchronous reset assertion and synchronous reset release.
A hybrid workflow can retrieve lexical and semantic candidates and combine their ranks, then inspect the passages. Do not simply compare their raw scores: the scales and meanings differ by retrieval method. A reranker is another selection stage that can reorder candidate passages for a question; it still needs evaluation. Keep the version and access filters independent of ranking so a high score cannot make an unauthorized or obsolete passage eligible.
The approved collection may lack the requirement, or the text may express it with different words. Inspect the versioned source and extraction coverage, then try a controlled synonym or a reviewed alternative retrieval method. If the source truly lacks the fact, changing retrieval cannot justify inventing an answer.
Create a small review set before choosing more machinery: an exact identifier, a paraphrase, a wrong-version trap, an ambiguous question, and an unsupported question. For each one, label the acceptable source chunks and expected abstention behavior. Inspect whether the right evidence appears among the candidates and whether irrelevant candidates are being accepted. A toy keyword score is not a calibrated confidence value, and this course supplies no measurements comparing retrieval approaches.
Keyword matching cannot interpret units, negation, or implied timing relationships. Semantic matching can return plausible neighbors with incompatible conditions. References: SQLite FTS5 for lexical indexing concepts and Sentence Transformers semantic search for query/document retrieval concepts, both consulted 2026-10-18. No external search library or embedding model is needed for the final lab.