Semantic repertory search, explained

Semantic repertory search retrieves rubrics by meaning instead of by matching words. Type "afraid when the lights go out" and a semantic engine surfaces "Mind; fear; dark, of" even though the query and the rubric share almost no vocabulary; a keyword engine, by contrast, finds only what contains the typed words. For an index written in compressed nineteenth-century clinical English — Kent's phrasing, Boenninghausen's abbreviations — this is the most consequential change in repertory access since digitisation itself.

Why keyword search was never enough

The repertory's vocabulary problem is structural. Patients say "I can't stand anyone feeling sorry for me"; Kent's index says "Mind; consolation; agg.". Patients say "burning feet at night, sticks them out of bed"; the index files sulphur-flavoured particulars under phrasing no patient has ever used. Print-era practitioners bridged the gap with years of page geography and cross-references; keyword search shortened the lookup but kept the burden — the practitioner still had to guess the index's words. A large share of mis-repertorization sits at exactly this translation step, which is why rubric-selection discipline has its own pedagogy.

How semantic retrieval works

A semantic engine converts every rubric in the corpus into an embedding — a numeric vector positioned so that texts with similar meaning sit near each other — and converts the practitioner's query into the same space. Retrieval is then a nearest-neighbour problem: return the rubrics whose vectors lie closest to the query's. The properties that matter clinically follow directly from the mechanism:

• It matches paraphrase and synonym: patient idiom reaches index idiom without translation.
• It is ranked, not boolean: results come ordered by closeness, and closeness is a model judgement, not a fact.
• It is corpus-bound: the engine retrieves only from the repertories it has embedded; it cannot invent a rubric, which distinguishes retrieval from generative suggestion — a distinction worth keeping sharp when vendors blur both into "AI search".
• It inherits the embedding model's training: a model weak on archaic clinical English will misplace exactly the rubrics that need the most help.

A well-built implementation lets the practitioner toggle between modes per query — semantic for meaning and context, keyword for exact words — and lets results be filtered and sorted, for example by remedy count, before a rubric is opened. The same semantic stack can extend to translated Complete Repertory sources and to materia-medica search across owned reference works.

The failure modes

Near-miss retrieval. The top result can be semantically adjacent but clinically wrong: "fear of the dark" and "fear; alone, of being; darkness agg." sit close in embedding space, and only reading the rubric and its sub-rubrics resolves which one the case supports. The mitigation is unchanged from print: open the rubric, read its scope and cross-references, then add.

False confidence in ranking. Rank order encodes vector distance, not clinical fitness. The third result may fit the case better than the first; treat the result list as a shortlist of candidates, not a verdict.

Repertory mismatch. Semantic engines embedded on one corpus behave differently on another. A query tuned by habit against Kent's phrasing will surface different candidates in Murphy's clinical chapters or the Complete Repertory's expanded rubric set; verify against the edition actually selected for the analysis.

Out-of-distribution queries. Highly abstract or theory-laden queries ("lacks grounding in the vital sphere") give embeddings nothing concrete to anchor on, and the engine returns plausible noise. Feed it what the repertory indexes: sensations, locations, modalities, concomitants, in plain language.

Using it well: a three-habit discipline

First, query in the patient's words, then verify in the index's words. The engine's job is the bridge; the practitioner's job is to confirm the far bank — rubric meaning, size, and grade — before the rubric enters the sheet.

Second, use keyword mode as the control: when a semantic result surprises, an exact-word search over the same corpus shows what the meaning-based ranking skipped. Running ten real patient phrasings through semantic and keyword modes side by side in any free repertory search is the cheapest possible audit of which mode actually reaches the rubric you would defend.

Third, keep the analysis arithmetic honest downstream — weighting and elimination amplify whatever the search step admitted.

Privacy belongs in the discipline too. Semantic queries describing a patient's symptoms are clinical data crossing a network; the vendor's stack and retention terms are therefore part of the evaluation. Reputable implementations document their search providers explicitly (Google and Pinecone are common in semantic stacks) and bind them with TLS 1.3 in transit, AES-256 at rest, and zero-retention agreements for any data routed through AI providers.

Verdict

Semantic repertory search removes the vocabulary tax that keyword search always charged. It repays the practitioner who verifies every retrieved rubric, and it quietly misleads the one who trusts the ranking. The mechanism is clear, the failure modes are predictable, and the discipline that contains them — patient's words in, index's words out, keyword mode as control — is the same discipline that has always separated clean repertorization from arithmetic dressed as analysis.

References

Kent, J. T. (1897) Repertory of the Homoeopathic Materia Medica, Lancaster: Examiner Printing House.

Kent, J. T. (1900) Lectures on Homoeopathic Philosophy, Chicago: Ehrhart & Karl.

van Zandvoort, R. (ed.) (1994 onward) Complete Repertory, IRHIS.

Murphy, R. (2024) MetaRepertory, Lotus Health Institute.

Similia (2026) "Searching for Rubrics (Semantic Search)", Similia Help Centre, https://similia.crisp.help/en/article/searching-for-rubrics-semantic-search-yhrokl/, fetched 2026-05-17.

Similia (2026) "Studying the Materia Medica", Similia Help Centre, https://similia.crisp.help/en/article/studying-the-materia-medica-1je9cgt/, fetched 2026-05-17.

Similia (2026) "Is my patient data secure?", Similia Help Centre, https://similia.crisp.help/en/article/is-my-patient-data-secure-sxmdfd/, fetched 2026-05-17.