Skip to main content

Lucy returns the evidence. Your model writes the answer.

Lucy RAG is a retrieval layer for financial AI. Ask a question and it returns the passages, tables, and XBRL facts from SEC filings that answer it, each linked to the filing and section it came from.

Request a pilotRead the docs
94.7%Recall@10 on FinanceBenchQdrant case study →+0.160Answer accuracy from retrieval alone, same modelQdrant case study →7 of 8Model–dataset pairs where Lucy matched or beat web searchBenchmark →
Your side
1Ask
A question comes in

From a person, an app, or an agent, in plain language or with filters.

Lucy
2Retrieve
Lucy searches the filings

Filtered by company, form, period, and Item, then ranked with dense and keyword search together.

Lucy
3Return
Evidence comes back

Passages, tables, and XBRL facts, each with its accession number and section.

Your side
4Answer
Your model writes it

Any LLM, grounded in that evidence and citing it. Lucy never writes the answer.

01 · Customer-facing financial AI

Ground your client-facing assistant in the filing

ForProduct managersML engineersBackend engineersBrokerages, wealth platforms, fintech

The problem

Your assistant is one wrong number away from a support ticket. Web search returns figures from the wrong period, and the model fills gaps with plausible ones.

What Lucy gives you

  • Evidence, not pages. The exact passages and tables behind a figure, with source links your UI can show.
  • Your model, your voice. Lucy supplies the grounding. Your LLM answers in your product's tone.
  • Current filings. New EDGAR filings are processed in near real time.

Pairs with these datasets

FINFinancial statementsSUMFiling summary

The request

What was Apple's total derivative fair value at the end of fiscal 2022?

Lucy returns

10-KApple Inc. · FY2022Item 8 · Notes · Derivative instruments0000320193-22-000108
Derivative assets       $7,136M
Derivative liabilities  $6,119M
As of September 24, 2022
chunk_type: table
10-KApple Inc. · FY2022Item 7A · Market risk0000320193-22-000108
“The Company uses derivative instruments, such as foreign currency forward and option contracts, to hedge certain exposures to fluctuations in foreign currency exchange rates…”
chunk_type: narrative
More chunks, ranked
→Your model answers $13,255 million in gross derivative fair value and cites both chunks. With web search context, the same model got the totals wrong.
Illustrative. Chunks follow the Lucy response format; excerpts are shortened.
+0.160Answer accuracy with Lucy context instead of web search, same model
0.952Answer accuracy with Lucy context, Claude Fable 5
Every chunkLinked to its filing, section, and accession number
Try the PlaygroundSource: Qdrant case study
02 · Investment research

Get the figure and the footnote behind it

ForEquity analystsCredit analystsResearch operationsAsset managers, brokers, independent research

The problem

A period-over-period comparison means two 10-Ks, two tax notes, and an MD&A paragraph, found by hand. A copilot built on web search gets the direction right and the number wrong.

What Lucy gives you

  • XBRL facts, not just prose. Financial statement questions return tagged values with their period context.
  • The note that explains it. The narrative from the tax note or MD&A comes back alongside the figure.
  • One filing, no mixed periods. Company, form, and period filters pin the exact filing.

Pairs with these datasets

FINFinancial statementsREVIncome statement flowCAPDebt instruments

The request

Corning's effective tax rate, FY2021 vs. FY2022, and what drove the change.

Lucy returns

10-KCorning Inc. · FY2022XBRL · EffectiveIncomeTaxRateContinuingOperations
FY2022  22.9%
FY2021  20.2%
chunk_type: xbrl
10-KCorning Inc. · FY2022Item 8 · Notes · Income taxes
The effective rate rose on a net provision for changes in tax reserves, differences arising from foreign earnings, and valuation allowance assessments.
chunk_type: narrative
More chunks, ranked
→Your copilot writes 20.2% → 22.9%, up 2.7 points, explains the drivers, and links each claim to its chunk. With web search context, it reported a decline.
Illustrative. Chunks follow the Lucy response format; excerpts are paraphrased and shortened.
94.7%Recall@10 on FinanceBench
0.774MRR on FinanceBench
XBRLTagged facts returned with their period context
Request a pilotSource: Qdrant case study
03 · Filing search and pipelines

Query SEC filings at scale, with real filters

ForData engineersQuant developersPlatform engineersResearch platforms, data vendors, quant funds

The problem

Keyword search over EDGAR misses tables and XBRL facts, and returns whole documents with no sense of which form, period, or section a match came from.

What Lucy gives you

  • Structured filters. Company, form type, date range, Item, and accession number are indexed fields.
  • Hybrid retrieval. Dense and keyword search fused into one ranking, over section-aware chunks.
  • Broad SEC coverage. Corporate, ownership, and fund filings, including 10-K, 10-Q, 8-K, DEF 14A, Forms 3, 4, and 5, 13F, and 13D/G, from 2016 to today.

Pairs with these datasets

All 17 datasets, through the same API

The request

{ "query": "chief executive officer transition",
  "form": "8-K", "item": "5.02",
  "date_from": "2025-01-01", "limit": 50 }

Lucy returns

{ "form": "8-K",
  "company": "…", "filing_date": "2025-…",
  "item": "5.02", "accession_number": "…",
  "chunk_type": "narrative", "text": "…" }
… one row per match, each with its chunk text and a link to the filing
→Every match carries its accession number, Item, and chunk text, so results drop straight into a pipeline or an index.
Illustrative request and response shape. See the Docs for the exact parameters.
2016–todayAbout ten years of SEC filings
HybridDense and keyword search, fused into one ranking
Filters firstApplied inside the query, before ranking
Read the docs
04 · AI agents

Give your agent a tool that returns citable context

ForAgent developersLLM platform engineersSolutions architectsAny agent framework, through MCP or REST

The problem

Agents need a tool that returns structured, citable context. A page of HTML to parse on every call is not that tool, and a summary with no source is worse.

What Lucy gives you

  • MCP server and REST API. Register Lucy as a tool, and the agent decides when to call it.
  • Provenance on every chunk. Accession number, Item, and chunk type come back with the text, ready to cite.
  • Works with smaller models. An open-weight Gemma model with Lucy context outscored frontier models using web search.

Pairs with these datasets

TLFiling timelineSUMFiling summary

The request

tools/call  lucy.search
{ "query": "changes to risk factors", "company": "NVIDIA",
  "form": "10-Q", "item": "1A", "latest": true }

Lucy returns · top result

{ "company": "NVIDIA Corporation", "form": "10-Q",
  "item": "1A", "chunk_type": "narrative",
  "accession_number": "…", "period_end": "…",
  "text": "…",
  "source": { "url": "https://www.sec.gov/…" } }
→The agent writes its summary and attaches one accession number per claim. Retrieval and citations come from Lucy; the reasoning is the agent's.
Illustrative request and response shape. See the Docs for the exact tool definition.
0.863Answer accuracy, open-weight Gemma model with Lucy context
MCP + RESTSame payload either way
CitableAccession number and Item on every chunk
Connect through MCPSource: LucyRAG benchmark
05 · Compliance and legal

Show the source for every statement

ForCompliance officersLegal counselInternal auditRegulated firms, audit and advisory

The problem

A figure nobody can trace is a regulatory risk. Every statement in a memo needs the filing and the section it came from.

What Lucy gives you

  • Chunk-level traceability. Each passage points to its filing, Item, and accession number.
  • Documents incorporated by reference. When a 10-K points to the proxy statement, Lucy retrieves the DEF 14A too.
  • Scoped retrieval. Restrict a search to one filing or a date range when policy requires it.

Pairs with these datasets

RISKLegal proceedingsINSDInsider transactionsOWNBeneficial ownership

The request

Which legal proceedings did Apple disclose in fiscal 2022? Give me the filing reference for each.

Lucy returns

10-KApple Inc. · FY2022Item 3 · Legal proceedings0000320193-22-000108
The Epic Games antitrust lawsuit over App Store operations, filed in the U.S. District Court for the Northern District of California, and its status on appeal.
chunk_type: narrative
10-KApple Inc. · FY2022Item 1A · Risk factors · Legal and regulatory0000320193-22-000108
Ongoing antitrust investigations and litigation in several jurisdictions, and the risks they pose to the business.
chunk_type: narrative
Every matching chunk in the period, in filing order
→Your model drafts a memo that lists each matter with its filing reference, and a reviewer can open every chunk behind it.
Illustrative. Chunks follow the Lucy response format; excerpts are paraphrased.
Source-linkedEvery chunk, every time
By referenceProxy statements cited in a 10-K are retrieved too
ScopedLimit a search to one filing or a date range
Talk to sales
Questions in plain language

Lucy RAG

Send a question, get back the passages, tables, and XBRL facts that answer it. Best when your model writes the answer.

Read the docs
Data, field by field

Structured datasets

17 datasets, from properties to insider transactions, with evidence on every value. Best when you need clean fields in a table or a model.

Browse the Data Catalog

Bring your team's questions. We'll run them on your filings.

A pilot runs Lucy RAG on the companies and forms your team covers and returns the evidence for review before any integration work.

Request a pilotTry the Playground