bilbysearch

This repo is a work in progress - it is a package for dedicated search code in future search pipelines at Bilby.

bilbysearch turns articles into vectors - by default with intfloat/multilingual-e5-large-instruct, or with any of the 31 other models in its registry, locally or through OpenRouter - and reads and writes the daily parquet files those vectors live in. It searches those articles by meaning (vector search) or by keyword (BM25, in English or Chinese), and reranks the results with Jev.

It exists so each model's contract lives in one place. What a model expects - its output width, context window, query instruction and document prefix - is recorded once, in model_config.py, and every function takes the model by name and applies that record. Which model production uses, and pipeline rules such as the title+body join, are in production_config.py. So notebooks, batch backfills and retrieval pipelines cannot drift apart.

Install

pip install bilbysearch

That one command brings embedding, storage and search: the model stack (sentence-transformers, torch), gs:// support (fsspec, gcsfs, google-cloud-storage) and BM25 keyword search (bm25s, and jieba to split Chinese into words). torch is the heavy part, at around 2 GB once CUDA wheels are involved.

Reranking with Jev is optional, because it brings bilbystats and its LLM client libraries:

pip install "bilbysearch[rerank]"

[all] installs everything. The [embed] and [gcs] extras from 0.1.x are still accepted, so an existing "bilbysearch[embed,gcs]" dependency keeps working, but they no longer add anything.

Usage

import bilbysearch

# Embed a frame of articles. Rows with no text get a null vector rather than a
# meaningless one, and three provenance columns record how the vectors were made.
articles = bilbysearch.read_parquet("daily_2026-01-31.parquet")
embedded = bilbysearch.embed_dataframe(articles)
bilbysearch.write_parquet(embedded, "daily_2026-01-31_embedding.parquet")

Queries and documents are embedded by different functions on purpose: the production model expects an instruction prefix on queries and none on documents, and skipping it measurably degrades retrieval. embed_query adds the query instruction and embed_texts the document prefix, so neither can be forgotten or put on the wrong side.

query = bilbysearch.embed_query("Chinese monetary policy easing")
passages = bilbysearch.embed_texts(["The PBoC cut the reserve requirement ratio."])
similarity = query @ passages[0]  # vectors are L2-normalised, so this is cosine

The model is loaded on first use and kept loaded - the two most recently used models stay in memory - so embedding a frame and then a query loads it once. clear_embedder_cache() releases them. Every function also takes device ("cuda", "mps" or "cpu"; the default lets sentence-transformers choose).

Choosing a model

Every function takes the model's name as model_name, defaulting to the production model. The name is all you pass: its width, context window and prefixes come from its entry in the registry.

bilbysearch.list_local_embedding_models(language="zh")  # "en", "zh" or "multilingual"
bilbysearch.get_model_spec("BAAI/bge-m3")                # what the registry records

embedded = bilbysearch.embed_dataframe(articles, model_name="intfloat/e5-large-v2")
query = bilbysearch.embed_query("PBoC easing", "intfloat/e5-large-v2")  # gets "query: "

That dataframe's articles were embedded with e5-large-v2's "passage: " prefix, and its embedding_model column says intfloat/e5-large-v2: the one name picks the model and labels the output, so the two cannot disagree. Embed queries with the same model as the documents; vectors from two models are not comparable.

Mistakes fail before anything is loaded or sent:

• An unknown name is refused, with the nearest known names suggested ("baai/bge-m3" -> did you mean 'BAAI/bge-m3'?).
• A text that already starts with its model's prefix is refused, e.g. "passage: ..." for e5-large-v2, which would otherwise be embedded as "passage: passage: ...". Pass texts without prefixes.
• A loaded model that disagrees with its registry width is refused, so its vectors never reach a vector store sized from the registry.

To switch production to another model, change MODEL_NAME in production_config.py to a local model from the registry; its width, context window and prefixes follow. tests/test_production_config.py pins the production values, so update it in the same change. Existing vectors must be re-embedded, and a vector store of a different width recreated.

To add a model, add its entry to model_config.py: its output width (measured from real output, not read from its card), context window, query instruction and document prefix (copied character for character from its card), language, whether it needs trust_remote_code, and where it runs.

Through OpenRouter

The same steps can run on OpenRouter's servers instead of this machine, for the models the registry records as hosted there (list_openrouter_embedding_models(); the production model is not one of them). Set OPENROUTER_API_KEY, then pass the model's registry name - the id OpenRouter uses and the model's prefixes come from its entry:

model_name = "intfloat/multilingual-e5-large"

embedded = bilbysearch.embed_dataframe_openrouter(articles, model_name=model_name)
passages = bilbysearch.embed_texts_openrouter(["The PBoC cut the RRR."], model_name)
query = bilbysearch.embed_query_openrouter("PBoC easing", model_name)

Locally, a text longer than the model's context window is cut short: e5-large reads only its first 512 tokens. OpenRouter would receive the whole text, so where the registry names the model's tokenizer (e5-large and the Perplexity models), each text is cut to the same window with that tokenizer before it is sent, and OpenRouter embeds exactly what the local model reads. OpenAI's and Gemini's tokenizers are not public, so their texts are sent whole.

Before mixing OpenRouter and local vectors in one index, check they match:

report = bilbysearch.compare_embeddings(local_vectors, openrouter_vectors)
report["all_match"]  # True if every pair has cosine similarity >= 0.99

check_openrouter_matches_local.py runs this check for multilingual-e5-large.

Threading

embed_texts_openrouter and embed_dataframe_openrouter send texts in batches of batch_size (default 32) per request, and use threads to keep up to max_concurrent_requests (default 8) requests in flight at once. Vectors come back in input order. Lower max_concurrent_requests if OpenRouter rate-limits you; 1 sends one request at a time.

embedded = bilbysearch.embed_dataframe_openrouter(
    articles, model_name=model_name, max_concurrent_requests=4
)

For async code, embed_texts_openrouter_async and embed_query_openrouter_async take the same arguments and can be awaited without blocking the event loop:

query = await bilbysearch.embed_query_openrouter_async("PBoC easing", model_name)

Upgrading from 0.2

0.3 takes models by name. Code written for 0.2 fails with an error saying what to pass instead, rather than running with the wrong settings:

Searching

vector_search scores every article in an embedded frame against a query and returns the closest, best first, with a cosine_similarity column. To search stored days, read them into one frame first:

uris = [bilbysearch.parquet_uri(prefix, date, "_embedding") for date in dates]
embedded = pd.concat([bilbysearch.read_parquet(uri) for uri in uris], ignore_index=True)
candidates = bilbysearch.vector_search("Chinese monetary policy easing", embedded, top_k=100)

The query is embedded with model_name - the production model unless you say otherwise - and it has to be the model that made the article vectors. Frames from embed_dataframe record that model in their embedding_model column, so a search with a different one, or over a frame that mixes days embedded by two models, is refused rather than returning a meaningless ranking.

It is brute force: every vector is loaded and compared in memory. That is exact and fine for weeks of articles; a corpus too big to load at once needs a vector store.

Keyword search with BM25

vector_search matches meanings; bm25_search matches words. It finds exact names, codes and figures that an embedding can blur, and misses the paraphrases an embedding catches. It needs no embeddings, so it runs on the raw daily files. Build an index once, then search it as often as needed:

day = bilbysearch.read_parquet("daily_2026-01-31.parquet")
index = bilbysearch.build_bm25_index(day, language="zh")  # title+body, split with jieba
hits = bilbysearch.bm25_search("央行 降准", index)         # milliseconds per search
hits[["title", "bm25_score"]]

Choose the language and the columns together. A query only matches words in its own language, and each language is split into words differently. In the daily files title+body is mostly Chinese and title_en+body_en is English:

en_index = bilbysearch.build_bm25_index(day, language="en", text_col="title_en+body_en")
hits = bilbysearch.bm25_search("reserve requirement cut", en_index)

• "en" splits text the way bm25s does by default - lowercased words of two or more letters or digits, English stopwords removed - so results match the team's English BM25 baseline.
• "zh" splits Chinese into words with jieba's search mode, which also emits the shorter words inside long ones: 中国人民银行 gives 中国, 人民, 银行 and the whole name, so a query for any of them finds it. Full-width letters and digits (ＧＤＰ, ４.８％) read as GDP and 4.8%, and Latin terms in Chinese text (PBoC, 5G) are kept.

language has no default, and an English index refuses a query with Chinese in it rather than silently matching nothing.

What comes back. Up to top_k (default 100) articles that contain at least one of the query's words, best first, with a bm25_score column - ready for rerank_jev. Articles with none of the words are left out, so fewer than top_k can come back. A score depends on the whole index as well as the article - how rare each word is across it, how long its articles are on average - so it ranks articles within one search. It is not comparable across indexes, and is not a relevance cutoff.

Tuning. build_bm25_index takes BM25's two parameters. k1 (default 1.5) sets how much repeating a word raises a score: higher lets each repeat keep adding, lower makes a word count about the same however often it appears. b (default 0.75) sets how much long articles are discounted, from 0 (not at all) to 1 (in proportion to their length). bm25s works out each word's contribution to each article's score when the index is built, which is what makes searches fast, so k1 and b are fixed for an index: build another to try other values.

Cost. Building reads every article: 2-3 s on a laptop for a day's Chinese title+body (around 1,300 articles), and 0.3 s for its English translations. Each search after that takes milliseconds. The index lives in memory.

Known limits of "zh":

• jieba's dictionary is simplified Chinese, so traditional text is split poorly.
• A name the dictionary does not know is split into pieces - 宁德时代 gives 宁德 and 时代 - which match more loosely than the whole name would.
• Search mode trades some precision for recall: 人民币 (the yuan) also gives 人民, which is in 中国人民银行 (the People's Bank of China), so a query about the yuan also finds People's Bank articles, lower down.

ex_bm25_search.py runs these searches on a few articles.

Reranking with Jev

Cosine similarity finds articles that are about the same things as the query, and BM25 finds articles that use the same words. rerank_jev asks TypeSafe's Jev whether each candidate is actually relevant, and re-sorts by its calibrated probability. It needs the [rerank] extra (see Install). Pass it the top 50 to 100 candidates from vector_search or bm25_search; the call goes through bilbystats.batch_jev_api, routed through OpenRouter by default, so it needs an OPENROUTER_API_KEY in the environment. Pass provider="typesafe" or provider="vercel" to route elsewhere, each with its own key.

candidates = bilbysearch.vector_search("Chinese monetary policy easing", embedded)
reranked = bilbysearch.rerank_jev("Chinese monetary policy easing", candidates, min_relevance=0.5)
reranked[["title", "jev_relevance", "jev_confidence", "jev_probabilities"]]

jev_relevance is the probability that the article is relevant, jev_confidence is Jev's confidence in that answer and jev_probabilities is the full yes/no dict. Because the probability is absolute, min_relevance works as a cutoff as well as a rank; ties keep the incoming order.

Development

uv sync --all-extras
uv run pytest
uv run ruff check . && uv run ruff format .

The tests use stub models and a faked OpenRouter, so the suite runs in a couple of seconds, never downloads a real model and never makes an API call.

Licence

MIT — see LICENSE.