Natural-language homeopathic search

Natural-language homeopathic search lets you type a symptom phrase the way the patient said it — "fear of the dark", "headache better lying down" — and get back a ranked list of repertory rubrics, materia-medica passages, and remedy hypotheses to test against the case. The interface is now standard across modern homeopathic software, sitting on top of vector-embedding retrieval.

How it works

There are four mechanical stages.

The input phrase is normalised and embedded into a high-dimensional vector by a sentence-embedding model — current vendors typically use OpenAI, Cohere, or open-source e5 and bge families behind a vector database such as Pinecone, Weaviate, or pgvector. Every rubric and materia-medica passage in the corpus has been embedded under the same model at index time, with rubric path and source attribution preserved. An approximate-nearest-neighbour query returns the top-k candidates ranked by cosine similarity. A re-ranking pass — often a cross-encoder or a small language model — filters and re-orders so that semantically near but clinically wrong matches drop down the list.

What you see is a ranked candidate set, not a prescription. A responsible implementation surfaces the rubric path verbatim (for example, MIND; FEAR; dark, of) so you can read it in its original repertory context before adding it to a repertorisation. The candidate set feeds your judgement; materia-medica confirmation, modalities cross-check, and the constitutional reading still do the load-bearing work.

Two structural choices separate a serious implementation from a marketing demo. The first is corpus. A vendor that has embedded Kent's Repertory, Boenninghausen's Therapeutic Pocket Book, and a contemporary Complete Repertory edition with rubric paths intact returns a candidate set a Kentian or Boenninghausenian prescriber can act on; a vendor that has embedded only marketing copy and remedy summaries returns a candidate set that reads plausibly and prescribes poorly. The second is surfacing the cosine score and source pill on each result row, so you can distinguish a 0.91 match against an exact Kent rubric from a 0.62 match against a paraphrased modern textbook. Without those two surfaces the candidate set is opaque and you have no audit trail to defend the prescription against peer review.

Semantic search versus keyword search

Two search modes meet you in current software. They answer different questions and have different failure modes.

Reach for semantic search when the case notebook holds the patient's own language ("she keeps saying her head feels too heavy to hold up") and for keyword search when verifying a rubric path you already have in mind. The two are not interchangeable; a well-designed implementation surfaces both so you can cross-check without switching tools.

A defensive habit worth keeping: run the same case through both modes and compare the top five candidates. Where the two lists agree, the rubric is well-anchored in both the surface vocabulary of the repertory and the embedding model's notion of nearness; where they disagree, you have a flagged item to read more carefully before committing it to the chart. This is the same posture Boenninghausen and Kent already encoded — never trust a single rubric in isolation; verify against modalities and concomitants.

Inside a working stack

Two surfaces are typical. In the Repertory, semantic search accepts a free-text symptom — "fear of the dark", "headache better lying down" — and returns ranked rubrics drawn from the repertories the account has access to, including translated Complete Repertory sources where licensed. In the Materia Medica section, semantic search lets you enter one or more symptom phrases as separate entries and returns remedies ranked by how strongly they match across the selected classical and premium sources.

The technical stack matters. In Similia's case the semantic-search pipeline uses third-party providers including Google and Pinecone as part of the retrieval stack. Separately, for AI features that process clinical text and images — notes and photo analysis via OpenAI, live-audio transcription via Deepgram — Similia has signed Business Associate Agreements with both providers and uses zero-retention processing for covered AI data; submitted patient data is not used to train AI models. The BAA and zero-retention commitments apply to OpenAI and Deepgram as covered clinical-AI sub-processors; they do not extend to Google or Pinecone, which serve a different role. A natural-language search step is a covered-data event under most clinic privacy regimes, which makes a vendor's data-handling posture a prescribing-relevant fact, not a marketing footnote.

Two further capabilities are worth knowing about when evaluating any search interface. Premium materia-medica integration that draws on Vermeulen's Concordant Materia Medica and the Synoptic Materia Medica alongside the classical sources widens the corpus the embedding step has to choose from when a query is paraphrased rather than literal. A multi-language layer means a French- or Hindi-speaking prescriber can phrase a query in their working language and still get the rubric path returned in its canonical English form for citation.

Where the canon places this technique

Natural-language search is a retrieval interface, not a clinical claim. It does not change how prescribing works.

Hahnemann's Organon, Kent's Repertory and Lectures, Hering's Guiding Symptoms, Boenninghausen's Therapeutic Pocket Book, Boger's Synoptic Key, Vithoulkas's Science of Homeopathy, and Sankaran's sensation-method volumes describe a 200-year observational practice in which the rubric, the modality, and the constitutional reading are the load-bearing artefacts of a case. Inside that practice, natural-language search shortens the distance between hearing a patient phrase and finding the rubric that matches it. It is a clerical accelerant for an interpretive method — useful exactly to the degree it preserves the rubric's original wording and lets you audit the path back to the source.

The criteria that distinguish a usable implementation from a misleading one are concrete: rubric paths and source pills surfaced verbatim, semantic and keyword modes available side by side, and the third-party AI-provider and data-handling posture published in writing. A practitioner who wants to test the technique against their own case notebook can search the Similia repertory and judge the rubric-shortlist quality on real cases.

References

Hahnemann, S. Organon of Medicine, sixth edition; The Chronic Diseases.

Kent, J. T. Repertory of the Homoeopathic Materia Medica; Lectures on Homoeopathic Materia Medica.

Hering, C. Guiding Symptoms of Our Materia Medica.

von Boenninghausen, C. Therapeutic Pocket Book.

Boger, C. M. Synoptic Key of the Materia Medica.

Vithoulkas, G. The Science of Homeopathy.

Sankaran, R. The Substance of Homoeopathy and the sensation-method volumes.

Vermeulen, F. Concordant Materia Medica; Synoptic Materia Medica.

Similia product reference: Repertory Semantic Search and Materia Medica Semantic Search documentation; AI-provider and data-handling disclosure, fetched 2026-05-16. Is my patient data secure?, https://similia.crisp.help/en/article/is-my-patient-data-secure-sxmdfd/.