Materia medica search — semantic vs keyword
Materia medica search is a retrieval problem with a peculiar corpus: millions of words of nineteenth-century clinical English, organised remedy by remedy, where the same observation appears as "dread of solitude" in one author and "fear of being alone" in another. Two search models compete over this corpus. Keyword search matches the literal string; semantic search matches the meaning. Each handles the classical texts differently, fails in characteristic ways, and can be tested in five minutes at no cost.
The two models, side by side
| Dimension | Keyword search | Semantic search |
|---|---|---|
| Matching unit | Literal string, sometimes stemmed | Meaning, via vector representation of the passage |
| Vocabulary gap | Fails when the author's wording differs from yours | Bridges period vocabulary ("anguish", "dread") to modern phrasing |
| Precision | High when the term is distinctive (Ailanthus, Cicuta) | Can return thematically related but off-target passages |
| Auditability | Fully transparent — the string is there or not | Ranking rationale is opaque to the reader |
| Failure mode | Silent false negative: relevant passage simply not returned | Plausible false positive: confident-looking near-miss |
| Cost | Free across the public archives | Platform feature, often included on free tiers |
The most consequential difference is the failure mode. A keyword search that misses is invisible — you conclude the corpus is silent when only the spelling was wrong. A semantic search that misses is visible but seductive — it returns something fluent and adjacent, and you have to check the passage against the source. The first error wastes knowledge; the second demands discipline. Knowing which error your tool makes is most of search literacy in this domain.
Why the corpus punishes keyword search
The classical materia medica was written across a century by authors with different vocabularies, different translation chains (Hahnemann wrote in German; the standard English texts are Dudgeon's and Hempel's renderings), and different symptom taxonomies. Boericke compresses; Hering itemises; Kent narrates. The practical result: the recall of a literal-string search over this corpus depends on already knowing the period phrasing — which is precisely what a student does not yet know. The standard workaround, building synonym lists by hand, is the manual labour that semantic retrieval automates.
Keyword search keeps two real advantages. For distinctive technical tokens — remedy names, anatomical terms, Latin binomials — it is exact, fast, and auditable. And it never misrepresents relevance: the string is in the text or it is not. A competent workflow uses both models and knows which question belongs to which.
What semantic search actually does
Semantic retrieval represents each passage and each query as a point in a vector space, where proximity approximates similarity of meaning; results are ranked by that proximity rather than by string overlap. Over the materia medica this means a query phrased the way patients speak — "she bursts into tears when comforted" — surfaces the classical "weeping aggravated by consolation" passages without your having to guess the canonical wording.
The reading discipline it demands is source-checking. Because ranking is opaque, open the returned passage in its full context — author, source, surrounding text — before relying on it. A highlighted full-text view makes this check cheap. A passage without its source attribution is not yet ready for clinical use.
The five-minute test
Search quality claims are cheap; the corpus makes them testable. Take one real symptom phrase, in your own words — not a textbook phrase. Run it three ways: against a free full-text archive with keyword search, against a platform keyword search, and against a semantic search. Score each on two counts: did it return the classically expected remedies (for "weeps when consoled", the tradition indicates Pulsatilla and Natrum muriaticum prominently), and can you verify each hit's source in one click?
You can run the comparison today by pulling the same phrase through a free semantic repertory and materia medica alongside the open archives — your own phrase decides the outcome.
References
- Boericke, W. (1927) Pocket Manual of Homoeopathic Materia Medica with Repertory, ninth edition, Philadelphia: Boericke & Tafel.
- Hahnemann, S. (1811–1821) Materia Medica Pura, reprint edited and translated by R. E. Dudgeon (1880), London: Hahnemann Publishing.
- Hering, C. (1879–1891) The Guiding Symptoms of our Materia Medica, ten volumes, Philadelphia: F. E. Boericke.
- Kent, J. T. (1905) Lectures on Homoeopathic Materia Medica, Philadelphia: Boericke & Tafel.
- Manning, C. D., Raghavan, P. and Schütze, H. (2008) Introduction to Information Retrieval, Cambridge: Cambridge University Press.
- Médi-T (1998–2026) Homéopathe International — hypertext materia medica editions, http://www.homeoint.org/.
Verdict
Ready to act on this?