Query-Aware Compression Makes LLMs More Accurate on Financial Documents
Introducing bear-2-finance, now in private preview: tell it what you're asking and it keeps the part of the filing that matters. Models answer more accurately from its output than from the full document, at a third to half of the tokens.
Ask an LLM a question about a 10-K and almost none of what it reads is about that question. A filing runs 100,000 to 250,000 tokens, the largest in our test set 419,000, while the details that decide an analyst question sit in a few thousand of them. The rest is noise the model has to read past, and it degrades the answers.
bear-2-finance is a query-aware compression model optimized for financial data. Give it the document and what to focus on: a question, an instruction, or a sub-task from an agent. It keeps the most relevant fraction at a budget you choose, from 10% to 90%, in a single pass in under a second, and returns plain text your existing model consumes unchanged. The result is not just a smaller bill: models answer more accurately from the focused document than from the raw filing.
Because it conditions on the question, it drops into multi-hop question answering and analysis systems. A financial analysis rarely asks one question; when an agent breaks it into sub-questions, each step gets its own focused view of the filing instead of the full 200,000 tokens.
We measured this on FinanceBench: 150 analyst questions over full SEC filings, answered by Claude Haiku 4.5 and Gemini 3.5 Flash Lite, three runs per setting.
Accuracy holds down to a third of the filing
Figure 1: FinanceBench accuracy at compression settings that meet or beat each reader's uncompressed baseline (dashed grey), mean of three runs. Claude is Claude Haiku 4.5, scored on the 132 filings that fit its 200,000-token window; Gemini is Gemini 3.5 Flash Lite over all 150.
Both readers hold or beat their full-document baseline from 90% of the filing kept down to a third. Claude Haiku 4.5 reads half of each filing and scores 81.3 against 80.3 on the full document; Gemini 3.5 Flash Lite reads a third and scores 79.6 against 79.3. The focused document reads slightly better than the raw one.
Ahead of the pipelines teams actually run
A production pipeline has to answer every question, including the ones about the 419,000-token filing. Scored end to end on the same 150 questions, compressing every filing to half finishes 4.0 points ahead of a pipeline that truncates oversized documents to fit, and 9.3 points ahead of one with no overflow fallback, while sending half the tokens.
Figure 2: Four end-to-end pipeline policies, Claude reader, same 150 questions, identically scored. In the first bar a document that does not fit the window counts as a failure; 18 of the 150 filings exceed it.
Why not truncate, or reach for RAG?
Figure 3: Context selection at an equal token budget of 33% kept, Gemini reader, 150 questions. Every method sends the same number of tokens and differs only in which ones.
At an equal budget of a third of each filing, truncation loses 33 points and BM25 loses 10. A tuned embedding stack ties on accuracy, so the difference is operational: bear-2-finance produces a ready context in under a second on a filing it has never seen, while a retrieval stack is still chunking and embedding, sometimes for minutes on the largest documents. Up to 20x faster, with no embeddings to pay for, no vector store to run, and no index to keep fresh. One stateless call on the document you already have.
Cost
Two to three times fewer input tokens cuts the reader bill and prefill latency by the same factor. Each question pays one fast small-model pass in exchange for a much smaller large-model read, and in a multi-hop system those savings repeat on every step. Or spend them on a better reader: at a third of the tokens, Gemini 3.5 Flash scores 84.4, 5.1 points above Flash Lite reading everything, for half the cost of the full filing on Flash.
Try it on your filings
bear-2-finance is in private preview. Start at 50% kept, confirm on your own questions, and turn the dial down.
Get access to the private preview
One endpoint, one budget parameter, no index to build. bear-2-finance is rolling out to design partners now.