Lucy RAG · Retrieval layer for financial AI
Filings are where financial analysis starts.
SEC filings are the primary source for a public company's risks, results, business structure, and executive pay. Lucy RAG is a retrieval layer for financial AI, built around how filings are actually structured, and it plugs into your AI assistant via MCP. Lucy returns the evidence. Your model writes the answer.
Every Lucy license includes the MCP server. Compare enterprise licenses and API plans.
01 · Why Lucy
Lucy treats filings as filings.
Filings are long, dense with tables and footnotes, and structured differently from one form to the next. Generic document search flattens them until the numbers lose their meaning, and when a model can't find the filing, it answers from memory. Lucy is a retrieval layer built for the structure of SEC filings from day one, so every chunk traces back to the official filing.
Form-aware retrieval
Lucy figures out which company and which form to look at, then searches within that form's structure, down to the Item.
Tables and footnotes stay intact
Numbers come back tagged with their period and line item, not stripped of context.
Filings are pre-processed
Key facts, links to prior filings, and structured data like customer concentration ship alongside the passages.
Every answer is source-linked
Each chunk carries its accession number and Item, ready to cite. If the filings don't say, Lucy says so.
02 · Where Lucy fits
Lucy is the retrieval layer. Generation is yours.
Between a question and its answer there are two jobs: finding the right evidence in the filings, and writing the answer from it. Lucy owns the first one end to end. It parses the question, figures out which company and which forms to search, and returns the top-ranked chunks that answer it: passages, tables, and structured facts, each linked to the filing and section it came from. The second job is entirely yours, with no constraints from us. Use the chunks as they land in ChatGPT or Claude, or pipe them into your own stack and let whatever model you choose write the answer.
A question comes in
From a person, an app, or an agent, in plain language.
Lucy finds the evidence
- Resolves the company and the forms to search
- Searches the filings with structure intact
- Returns the top-ranked chunks, each with its accession number and section
Your model writes the answer
Any LLM, in your own stack, grounded in that evidence and citing it. Lucy never writes the answer.
Evidence with provenance
Passages, tables, and structured facts, each traceable to the filing and section it came from.
Keep the pipeline you built
Your existing RAG stack stays as is. Add Lucy as one more retrieval source via MCP or the REST API, and keep your reranking, filters, and prompts.
Bring your own model
ChatGPT, Claude, Gemini, an open-weight model, or your own fine-tune. Grounded in Lucy's chunks, even a smaller model holds its own.
How much the context changes answers: Lucy RAG benchmark results on FinQABench, FinanceBench, and recent financial statements.
03 · Connect
Connect from the client you already use.
Lucy is a remote MCP server. Register it as a tool, and every client signs in through the same OAuth flow. Integrating without MCP? See the developer docs for the REST API and dataset schemas.
https://lucy-rag-mcp.ezar.co.kr/mcpChatGPT
web · desktop · mobile- Open Lucy — SEC Filings Search in the ChatGPT app directory and click Connect.
- Sign in at
login.lucydata.aiwith your Lucy account. - Ask: “How does Apple make money?” If ChatGPT doesn't call Lucy on its own, say “Use Lucy to find Nvidia's business segments.”
On Enterprise or Edu workspaces, an admin may need to allow the app first.
Claude
claude.ai · desktop · mobile- Open Lucy — SEC Filings Search in the Connectors Directory and click Connect.
- Sign in at
login.lucydata.aiwith your Lucy account. - In the tools menu next to the message box, make sure Lucy is toggled on, then ask: “What are Apple's main risk factors?”
The first call may prompt for permission. Pick Always allow if you plan to use Lucy regularly.
Codex
CLI · desktop- Add Lucy to
~/.codex/config.tomlas an MCP server, using the URL above and the OAuth settings we send you once access is enabled. - Launch Codex. It opens the sign-in page in your browser; sign in with your Lucy account.
- Ask about a company. If Codex doesn't call Lucy, name it: “Search lucy-rag for …”.
Requires Codex 0.158 or later.
Claude Code
CLIRegister Lucy with the OAuth flags we send you once access is enabled:
claude mcp add --transport http lucy-rag \ https://lucy-rag-mcp.ezar.co.kr/mcp
- Run
/mcpand sign in with your Lucy account. - Ask about a company. Claude Code calls
rag_contextand cites the passages.
Custom MCP client
Streamable HTTP · OAuth 2.1- Point your client at
https://lucy-rag-mcp.ezar.co.kr/mcp. Authorization server metadata is published at/.well-known/oauth-protected-resource. - Send us your client's OAuth callback URL so we can register it, and use the OAuth client settings we send back.
- Call the
rag_contexttool with a question. Reuse the returnedthread_idfor follow-ups.
There's also a REST API for integrations without MCP. Read the docs.
04 · Example questions
Name the company. Ask in plain language.
You don't need to name a form or an Item; Lucy figures out where to look. Every question below returns filing evidence from the production server.
| Use case | Example question | Where Lucy looked |
|---|---|---|
| Business model | How does Apple make money? | 10-K · 10-Q · 8-K |
| Competitive position | What competitive advantages does Amazon have over its competitors? | 10-K · 10-Q · 8-K |
| Customers | Who are CoreWeave's biggest customers? | 10-K · DEF 14A |
| Earnings drivers | Why did Nvidia's profit grow so much last year? | 10-K · financial statements |
| Risk factor review | What are Apple's main risk factors in their 2024 10-K? | 10-K · Item 1A |
| Foreign issuer | How does Toyota describe its electrification strategy in its latest 20-F? | 20-F · 6-K |
| Canadian issuer | What risk factors does Shopify disclose in its latest 40-F? | 40-F · 6-K · 8-K |
| Monthly revenue | What monthly revenue did TSMC report in its recent 6-K filings? | 6-K |
| Governance and pay | Who are Tesla's named executive officers and how is CEO pay structured? | DEF 14A · 10-K |
| Comparison | Compare Apple's and Microsoft's discussion of AI-related risks in their 10-K filings. | 10-K · 8-K |
- Follow up in the same conversation. After the Apple risk-factor question, ask "How did those risks change year over year?" and Lucy keeps the context.
- Always name the company. Adding a form or topic like 10-K, proxy statement, or risk factors sharpens the match.
- If the assistant answers from memory instead of calling Lucy, say so: "Use Lucy to check Tesla's R&D spend."
- Out of scope: today's stock price, news, private companies, or companies that don't file with the SEC. Lucy returns no chunks, and the assistant tells you the filings don't have an answer.
05 · Privacy
Only what a search needs.
Lucy receives your question, an optional conversation ID and the OAuth access token that proves you're signed in. It returns passages from public SEC filings. Service logs (request metadata, the signed-in account ID and error codes) are retained for 30 days for operations and abuse prevention. Hosting and sign-in run on AWS; the retrieval layer sends the query text to the OpenAI API for query planning and embeddings. We don't train on your data, and we don't sell it.