FDRE
View source

About FDRE

Research infrastructure that shows its work

FDRE converts SEC filings into auditable retrieval results, structured facts, point-in-time feature data, and reproducible event-study inputs for research teams.

Research and data engineeringQuant research engineeringNo trading-strategy claims

See it work

The research console, replayed

Move through cited answers, point-in-time retrieval, cross-sectional screening, and signal studies. Pick a mode or let the replay advance.

Cited answers from filings

Ask

Retrieve, rerank, and verify filing evidence before answering.

Ask
AAPL · text Supply-chain changesMETA · earnings Latest quarterNo forecasts Unsupported request

Open the live console →

Verified production corpus

Measured scale, not projected scale

These counts are read live from the production database. The S&P 500 universe uses current constituents and is therefore survivorship-biased.

498 / 499
S&P 500 primary tickers indexed
2,838
SEC filings parsed and chunked
2,761,016
Chunks with stored embeddings
512
Voyage embedding dimensions

Specimen outputs

What a single result looks like

A typed financial fact resolved to its filing, and the bounded retrieval run that produced a verified answer: the artifacts behind every console response.

SEC 10-QMETA
Net income$26.77BThree months ended March 31, 2026
Revenue
$56.31B
Diluted EPS
$10.44

Accepted April 30, 2026

Retrieval runVerified
  1. 01
    Resolve issuerMETA only
  2. 02
    Hybrid retrieveText · tables · facts
  3. 03
    Verify citation100% text overlap
Top rerank score0.648

Engineering evidence

Built for reproducible research

The public service is research infrastructure, not a low-latency trading system or a portfolio backtest.

Indexed retrieval

PostgreSQL GIN full-text search and float16 HNSW cosine search over pgvector.

Point-in-time controls

SEC acceptance timestamps, availability boundaries, amendments, and as-of filtering.

Structured facts

Raw Company Facts plus canonical revenue, margins, cash flow, debt, and EPS.

Filing differences

Comparable periods with added, removed, and materially changed passages.

Research panel

Versioned JSON, CSV, and Parquet issuer-period features with leakage checks.

Auditable operations

Ingestion manifests, recovery metrics, data-quality audits, and experiment IDs.

Research workflows

Six public demonstrations

  1. 01
    Single-name risk retrieval

    Find section-aware evidence and inspect every retrieval score and citation.

  2. 02
    Table and XBRL extraction

    Query typed financial facts with the linked filing and narrative evidence.

  3. 03
    Filing change detection

    Compare the latest filing with its deterministic comparable period.

  4. 04
    Cross-sectional thematic research

    Scan broad themes while capping evidence per issuer for diversified results.

  5. 05
    Research datasets and event studies

    Build point-in-time issuer-period data and run benchmark-adjusted statistical tests.

  6. 06
    Published signal studies

    Four point-in-time filing studies covering disclosure, risk, composite, and earnings quality, with multiple-testing-adjusted inference and honest verdicts.

Methodology

Index offline, retrieve live

The bounded LangGraph retrieval workflow preprocesses the query, routes text, tables, and structured facts, reranks evidence, applies an evidence gate, and verifies every citation before returning an answer.

  1. Cached SEC ingest with acceptance timestamps
  2. Layout-aware text and table parsing
  3. Hybrid sparse and dense retrieval
  4. Typed facts, diffs, panels, and experiments
  5. Verified answer or deliberate abstention
Architecture