E-commerce search fails in two opposite ways. BM25 keyword search can't tell that "evening wristwatch" and "dress watch" mean the same thing. Pure vector search happily returns "something watch-like" when the shopper typed an exact SKU. This R&D engine runs both and fuses the results, on a Solr 9 core of about 540 real catalog products, without touching the production search path.
#1. The Pipeline
query ─▶ SKU short-circuit ─▶ split into item clauses ─▶ for each clause:
├─ price phrase parser (under / between / ~ / 12k)
├─ rule-based NLU (brand, category, colour, size, badge, type)
├─ fuzzy fallback (Solr ~2 edit distance on *_text fields)
├─ in-process ONNX embedding (all-MiniLM-L6-v2, 384 dims)
├─ BM25 keyword search (edismax)
├─ kNN vector search ({!knn f=vector topK=N})
└─ RRF fusion
─▶ merge clauses ─▶ de-dup by style ─▶ page ─▶ UI view model#2. Reciprocal Rank Fusion
BM25 scores and cosine similarities live on different scales, so you can't just add them. RRF ignores scores and uses only rank positions:
score(doc) = Σ over result lists 1 / (K + rank(doc)) K = 60A product ranked well by both retrievers rises to the top. One found by only one retriever still gets a fair share. The same primitive powers the "similar products" endpoint, where it fuses vector similarity with Solr's MoreLikeThis.
#3. Embeddings Without an API Call
- all-MiniLM-L6-v2 in ONNX format, run in-process with
Microsoft.ML.OnnxRuntime, tokenised withFastBertTokenizer. - Mean-pooled and L2-normalised into a 384-dimensional vector. The query is embedded live; product vectors are precomputed and stored in Solr.
- No network hop and no per-query cost for an external embedding API.
#4. Query Understanding Without an LLM
The NLU is deterministic and explainable: lexicon lookup and regex against a live vocabulary. A BackgroundService refreshes the real distinct values in the index (brands, colours, categories and so on) every five minutes using Solr facet queries. When the catalog changes, the parser updates without a redeploy.
- Typos: "Chrons" resolves via Solr's native fuzzy operator. A curated stop-word list (~60 words) stops common English words from being offered as fuzzy matches.
- Synonyms live in Solr, not in C# dictionaries. Gender and size moved to
*_textfields backed bysynonyms.txt. - Prices: "under X", "between X and Y", reversed phrasing, "~X" (±10%), "12,000" and "12k".
#5. Multi-Intent Queries
"red card case and gold tone watch under 5000" is two searches. Splitting on commas and a standalone "and" (while protecting "between X and Y" and "12,000") lets each clause run the entire pipeline with its own price and filters. Otherwise the watch clause's category filter would starve the card-case clause.
In "black and blue watches", "black" is only a modifier. Clauses that contain nothing but a colour, brand or gender word get merged into their neighbour, so they don't trigger a catalog-wide search for "black".
#6. The Details That Make It Feel Right
- Style-level de-dup: the same style in three colours shows once, with its colour swatches harvested through Solr collapse + expand.
- Phrase proximity (
pf2) reranks adjacent-word matches above bag-of-words hits. - Business signals (rating, sale count, availability) as an edismax boost function that nudges relevance but doesn't override it.
- Embedded SKUs ("watch CR6121") are forced to position one.
- Autosuggest through Solr's Suggester component for fast prefix and infix matches.
#7. Takeaways
- Hybrid beats either retriever alone, and RRF is the simplest fusion that works without score calibration.
- You don't need an LLM for query understanding in a bounded domain. A live vocabulary and good rules are faster, cheaper and explainable.
- Push linguistics (synonyms, fuzziness, suggesters) into the search engine, and keep the application code for orchestration.