VOXELL
NEWS
RELEASE // FORGE 2.0 RELEASE DISPATCH

Forge 2.0 Is Live: Two-Stage Retrieval, the Voxell Reranker, and 200-Question Signed Receipts

Jonathan Corners

Forge 2.0 ships two-stage retrieval with our in-house Voxell Reranker, 940,932 searchable public passages, and 200-question signed receipts across 4 corpora.

Forge 2.0 is live in production.

Most retrieval vendors publish a single aggregate score on a small academic dataset, hide the per-question results, and hope you never ask why their pipeline misses the exact clause you need inside a 200-page filing.

With Forge 2.0, we are shipping the two-stage retrieval engine we run in production, opening a zero-login public workbench across 940,932 real passages, and publishing the full 200-question retrieval receipts across four public corpora (7,817 documents, 980,885 total indexed passages, 800 scored questions) so you can inspect every question, every latency stage, and every miss yourself.

1. The Complete Forge 2.0 Receipt Breakdown

Across the four public corpora (800 scored questions total), the first result answers the question 83% of the time, one of the top three results answers it 90% of the time, and the right document is in the top ten 95% of the time.

Corpus Documents Passages nDCG@10 (95% CI) Answer@1 Answer@3 Doc@10 Recall@10 MRR@10 Public Receipt
SEC EDGAR Filings (10-K, 10-Q, 8-K) 2,010 183,955 0.841 (0.799 to 0.877) 91.0% 95.0% 97.0% 92.5% 0.812 Inspect SEC receipt
USPTO Patents (Granted full text & claims) 4,008 378,252 0.787 (0.742 to 0.834) 86.5% 91.5% 96.0% 90.0% 0.749 Inspect USPTO receipt
NASA Technical Reports (NTRS archive) 1,210 378,725 0.730 (0.677 to 0.784) 67.3% 78.4% 88.0% 81.5% 0.702 Inspect NASA receipt
arXiv Technical Papers (CS / AI preprints) 589 39,953 0.900 (0.869 to 0.929) 86.5% 96.0% 99.5% 98.5% 0.872 Inspect arXiv receipt
Total / Weighted Mean 7,817 980,885 0.815 82.8% 90.2% 95.1% 90.6% 0.784 All retrieval receipts

(Note: The three public-domain / open corpora, SEC EDGAR, USPTO Patents, and NASA NTRS, total 940,932 full-text passages and are live for interactive querying in the Forge 2.0 Public Workbench. For arXiv technical papers, we publish the questions and scores only because the prose belongs to its authors.)

2. What Moved the Number: Why the Embedding Model Alone Is Not the Bottleneck

Voxell’s Ingot-8B-R3 ranks #1 for English on the public MTEB leaderboard (English v2), with a 75.98 mean task score across 41 tasks. We have more reason than anyone to tell you that swapping in a stronger embedding model solves enterprise retrieval.

Our own receipts prove otherwise. On SEC filings, thousands of documents share identical regulatory form structures and accounting boilerplate. When we held the embedding model constant and measured the exact same 200 SEC questions at every architectural step, here is what actually moved nDCG@10:

Architectural Step (Same 200 SEC Questions) nDCG@10 Delta What Changed in the Pipeline
01. Cross-encoder reranker over flat chunks 0.467 Baseline A cross-encoder rereads the top dense candidates against the question and reorders them, but flat chunks lack filing identity.
02. Document identity in every passage 0.622 +0.155 Every passage carries a one-line identity header of the filing it came from: company name, SEC form, and reporting period.
03. Document-identity routing 0.843 +0.221 When a question names a company or filing, Stage 1 routes candidate retrieval to those documents first before widening.
04. Voxell Reranker (Forge 2.0 Production) 0.841 -0.002 (at ~1/5th compute) Our in-house Voxell Reranker replaced the larger general-purpose cross-encoder, matching 0.841 nDCG@10 (91.0% Answer@1, 95.0% Answer@3) at roughly one-fifth the compute per question.

The same pattern holds on USPTO Patents (0.670 baseline cross-encoder -> 0.801 with document identity and routing -> 0.787 with the Voxell Reranker at 1/5th the compute) and arXiv Papers (0.867 -> 0.917 -> 0.900).

3. Stage-1 and Stage-2 Latency Telemetry (No Hidden Clocks)

Single-number latency claims hide how a two-stage pipeline actually spends its time. In Forge 2.0, every query in the Public Workbench and API reports the stage split explicitly:

  1. Stage 1: GPU-Native Dense Candidate Sweep (87 ms P50 embedding latency) Embeds the incoming query in BF16 on dedicated GPU execution paths and sweeps the corpus to pull the top 50 candidate passages with document-identity routing. See our Forge Latency Methodology for the exact single-request harness.
  2. Stage 2: Voxell Reranker & Duplicate Collapse (~3.4 s over 50 long-context passages) Rereads all 50 candidate passages joint-encoded against the full question text, promotes the specific clause or table that answers the question over look-alike boilerplate, and collapses duplicate passages that appear across related filings or continuation patents.

4. Methodology Defense: How the Receipts Are Measured and Why We Publish the Misses

A retrieval benchmark is only useful if you know exactly what it measures and where its limits are:

  • 200 questions per corpus, locked baseline: Each corpus has 200 questions. Every question names its subject (the company, patent, report, or paper it is about) and has one specific passage labelled as the ground-truth source.
  • nDCG@10 with 95% bootstrap confidence intervals: The engine retrieves ten passages per question, and the receipt scores where the labelled passage landed (nDCG@10 from 0 to 1, Recall@1, Recall@10, MRR@10, and Doc@10). We compute a 95% bootstrap confidence interval on every corpus so variance is visible.
  • Independent reference-answer judging (Answer@1 and Answer@3): Because real documents often state the same material fact in more than one section (for example, an executive summary and a footnote), a calibrated model judge also reads each of the top three returned passages against the question’s reference answer and grades whether result #1 (Answer@1) or any of the top 3 (Answer@3) states the facts required to answer the question.
  • Why we state the caveat up front: A model wrote the test questions and reference answers from the source documents, which makes them easier than an adversarial human test set. They are not a comparison with any other vendor and not a promise about your private documents. They are a reproducible engineering yardstick for proving what moves retrieval quality on a locked baseline.
  • Every published receipt includes the misses: When you open the SEC Filings, USPTO Patents, NASA Reports, or arXiv Papers receipt pages, you see the individual questions and scores, including the questions the engine got wrong, alongside direct links to verify the source documents on SEC EDGAR, Google Patents, and NASA NTRS.

5. The Voxell Release Stream: What Ships Next (Forge 2.1 to 2.3)

Forge 2.0 kicks off our versioned release stream. Every release ships a measured capability directly into the product:

  • Forge 2.0 (Live Today): Two-Stage Retrieval, Voxell Reranker & Public Workbench Top-50 dense retrieval plus in-house neural reranking, duplicate passage collapse, signed 200-question receipts, and 940,932 public passages open at playground.voxell.ai.
  • Forge 2.1 (Fast Follow): Tabular Data & AST-Aware Code Ingestion Financial schedules, CSV/Parquet files, and multi-language source repositories break standard token chunkers. Forge 2.1 ships structure-preserving tabular ingestion (keeping column headers and row relationships bound to every cell) and syntax-tree-aware code chunking so functions, signatures, and call graphs remain intact.
  • Forge 2.2 (In Stream): Corpus Quality (VQS) & Intent Workspace Pre-flight Corpus Quality (VQS) contradiction and coverage scoring combined with our question-first Intent Workspace, surfacing conflicting policies and stale documents before a single user queries your Space.
  • Forge 2.3 (In Stream): Self-Serve Live Corpus Trials Drop a slice of your own PDFs, tables, or repositories into a private trial workspace, run your hardest queries with full stage-1 and stage-2 telemetry, and generate a scored retrieval baseline on your own data in minutes.

Try Forge 2.0 Now