796 pages of semantic search with no vector database, and what it cost to skip one

A static site with 796 indexed pages now runs meaning-based search entirely in the reader's browser, with no server, no vector database, no index service, and no model executing when you type ^[1]. That is worth attention because it is a working counterexample to the architecture most teams reach for without arguing about it: embed the corpus through an API, push the vectors into a hosted index, embed the query at request time, pay per call, and add a service that can be down ^[2].

The constraint came first, which is usually how this goes. According to the author's write-up on dev.to, the site has a house rule that no page may send a reader's input to a third party, and it is a static build on Cloudflare's static assets, so there was no search server to add and an embedding API call from the query box was forbidden by policy ^[3]. The other browser-side option, shipping a sentence transformer as ONNX or WASM, is tens of megabytes plus a warm-up delay and puts a model in the critical path of a text box ^[4].

The method is Model2Vec: run the transformer once at build time over a fixed vocabulary and keep only the per-word output vectors, which turns runtime embedding into a dictionary lookup and an average ^[5]. The table here was distilled from Xenova/all-MiniLM-L6-v2 over the corpus's own vocabulary, one contextless 384-dimensional vector per word, 8,900 words at present ^[6]. Each document vector is the L2-normalized mean of its words' vectors, quantized to int8 and base64 encoded ^[7]. The query is embedded by exactly the same rule in the same code, which is what keeps queries and documents in the same space ^[8].

What ships is three static JSON files from the site's own origin, gzip and brotli by the host, plus one inline module: 401 lines of vanilla JavaScript using only atob, Math, Map, Set and typed arrays, with no DOM, no Buffer and no imports ^[9][10]. Ranking is cosine against 796 unit vectors, which is 796 dot products of length 384 ^[11], roughly 305,000 multiply-accumulate operations per query ^[12]. The author reports that unpacking every document vector takes 12 ms in Node on his machine and a full three-channel query takes 4 to 10 ms ^[13].

The second-order property is the one operators should copy. Because the word table is the expensive artifact and the document vectors are cheap, a build-start hook re-embeds the live corpus from the table that already exists, with no model and no network; the model is only needed to mint vectors for genuinely new words ^[14].

That is also where the bill arrives. A static word table has no context, so word order is gone, "dog bites man" and "man bites dog" embed identically, and negation is invisible ^[15]. Coverage decays as the corpus grows: 4,689 distinct words in the live corpus currently have no vector, and 459 documents dropped at least one word when their mean was taken ^[16]. The index reports this about itself in its own header, with vocabCovers set to false ^[17]. Those 459 documents are about 58 percent of the 796 ^[18], and the uncovered word count is now more than half the size of the covered table ^[19].

Two things to watch. First, whether retrieval quality visibly degrades before someone reruns the distillation, since the header flags the condition but nothing forces the fix ^[17][14]. Second, how the approach behaves on a less distinctive corpus; this one is described as dense, cross-referenced and deliberately weird, which flatters averaged word vectors ^[20]. The numbers are self-reported and the search is live at artwaste.land/ask ^[21].