Skip to main content

​
Overview

SEC Filings is Finance Agent’s specialized retrieval system for extracting information from 10-K annual reports (2018-2025). When you need audited financial statements, executive compensation, risk factors, or detailed business descriptions, this feature routes your query to the official SEC documents that professional analysts rely on. Current Scope: 10-K filings only (annual reports). Support for 10-Q (quarterly) and 8-K (current events) is in development.

Planning-Driven Retrieval

Generates sub-questions and searches multiple sections in parallel for comprehensive coverage

Intelligent Section Routing

LLM automatically routes queries to relevant 10-K sections (Item 1, Item 7, Item 8, etc.)

Smart Table Selection

LLM selects relevant financial tables from income statements, balance sheets, and cash flow statements

91% Accuracy

Achieved 91% accuracy on FinanceBench (112 10-K questions), ~10s per question

​
How It Works

The SEC Filings agent is a dedicated retrieval agent optimized for structured 10-K documents. When the main agent determines your question requires 10-K data (via semantic routing), it invokes this specialized agent:
  1. Smart Planning - Breaks down your question into sub-questions and creates a search plan (table vs text queries)
  2. Parallel Retrieval - Executes all searches simultaneously (up to 6 workers)
    • Table searches: LLM selects relevant tables from financial statements
    • Text searches: Hybrid vector + keyword search with cross-encoder reranking
  3. Answer Generation - Synthesizes findings with [10K-N] citation markers
  4. Iterative Refinement - Evaluates quality (target 90%+) and replans if gaps exist (up to 5 iterations)
Why a separate agent? SEC 10-K filings have unique structure (15 sections, complex tables, financial statements) requiring specialized retrieval strategies that differ from earnings transcript search.

​
What You Can Ask

Best for: Audited annual figures, balance sheets, income statements, cash flow statements
  • “What was Apple’s total revenue in fiscal 2024?”
  • “Show me Microsoft’s balance sheet for 2023”
  • “What was Google’s cash and cash equivalents in 2024?”
  • “Compare Amazon’s total assets for 2023 and 2024”
Best for: CEO pay, stock awards, salaries, bonuses (ONLY available in 10-K, not in earnings transcripts)
  • “What was Tim Cook’s total compensation in 2023?”
  • “How much did Satya Nadella earn in 2024?”
  • “What stock awards did the CEO receive?”
  • “Show me executive compensation breakdown”
Best for: Material risks, regulatory issues, legal proceedings
  • “What are Apple’s main risk factors?”
  • “What risks did Meta disclose about regulation?”
  • “What legal proceedings is Tesla involved in?”
Best for: Business descriptions, segment breakdowns, workforce size, strategy
  • “How many employees does Microsoft have?”
  • “What are Amazon’s business segments?”
  • “What products does Apple sell?”
  • “Describe Google’s revenue streams”
Best for: Calculations requiring multiple line items
  • “What is Apple’s debt-to-equity ratio?”
  • “Calculate Microsoft’s gross margin”
  • “What is Amazon’s revenue per employee?”
  • “Show me R&D expense as a percentage of total operating expenses”

​
Query Examples

​
Executive Compensation

Question: What was Tim Cook's compensation in 2023?

Answer:
Tim Cook's total compensation for fiscal 2023 was **$63.2 million** [10K-1], consisting of:

| Component | Amount |
|-----------|--------|
| Base Salary | **$3.0M** [10K-1] |
| Stock Awards | **$47.0M** [10K-2] |
| Non-Equity Incentive | **$10.7M** [10K-1] |
| Other Compensation | **$2.5M** [10K-3] |

This represents a 12% decrease from 2022's **$71.9M** [10K-1].

​
Financial Statement Data

Question: What was Microsoft's total revenue in fiscal 2024?

Answer:
Microsoft reported total revenue of **$245.1 billion** for fiscal year 2024 [10K-1], 
up **16%** from fiscal 2023's **$211.9 billion** [10K-2].

**Revenue Breakdown** [10K-3]:
- Productivity and Business Processes: **$69.3B**
- Intelligent Cloud: **$96.8B**
- More Personal Computing: **$59.7B**

​
Derived Metric (Multi-Step Calculation)

Question: What is Apple's debt-to-equity ratio in 2024?

Answer:
Apple's debt-to-equity ratio for fiscal 2024 is **1.54**.

**Calculation:**
- Total Debt: **$106.6B** [10K-1] (long-term debt) + **$9.8B** [10K-2] (short-term) = **$116.4B**
- Total Equity: **$75.5B** [10K-3]
- Ratio: **$116.4B / $75.5B = 1.54**

This represents an increase from 2023's ratio of **1.41** [10K-4][10K-5].

​
SEC 10-K Section Guide

The agent automatically routes queries to the most relevant sections:
SectionContainsExample Queries
Item 1 - BusinessCompany description, products, operations, workforce size”How many employees?”, “What products does X sell?”, “Business segments”
Item 1A - Risk FactorsMaterial risks, uncertainties”What are the main risks?”, “Regulatory concerns”
Item 7 - MD&AManagement analysis, trends, performance discussion”Management commentary on growth”, “Trends”
Item 8 - Financial StatementsBalance sheet, income statement, cash flow, all financial line items”Total revenue”, “Cash”, “Assets”, “Liabilities”
Item 10 - DirectorsBoard members, executives”Who is on the board?”
Item 11 - Executive CompensationCEO pay, salaries, bonuses, stock awards”CEO compensation”, “Executive pay”
Important: Executive compensation data is ONLY available in 10-K filings (Item 11), never in earnings transcripts. Always use the SEC Filings feature for compensation queries.

​
Features

For complex queries, the agent decomposes your question into targeted sub-questions and creates a strategic search plan: Example: “What is revenue per employee?” 
{
  "sub_questions": [
    "What was total revenue?",
    "How many employees does the company have?"
  ],
  "search_plan": [
    {"query": "total revenue", "type": "table", "priority": 1},
    {"query": "subscription professional services revenue", "type": "table", "priority": 2},
    {"query": "number of employees", "type": "text", "priority": 1},
    {"query": "human capital workforce", "type": "text", "priority": 2}
  ]
}
Each query is optimized:
  • Short (2-5 words) for better embedding matches
  • Focused (one concept per query, not multiple crammed together)
  • No company names or years (already filtered by ticker and fiscal year)

​
Parallel Multi-Query Retrieval

All searches execute simultaneously:
  • Table queries: LLM selects top 5 relevant tables from available financial statements
  • Text queries: Hybrid vector + keyword search with cross-encoder reranking
  • Deduplicates and combines results

​
Smart Table Selection

Rather than embedding-based table search, the agent uses LLM intelligence:
  1. Fetches all tables for the ticker/fiscal year
  2. Creates table summaries: path | section | content preview
  3. LLM analyzes summaries and selects most relevant tables
  4. Returns selected tables with full content (tables are never truncated)

​
Iterative Quality Control

After generating an answer, the agent:
  1. Evaluates quality (0.0 to 1.0 scale)
  2. Identifies missing data points
  3. Replans with new search queries if quality < 90%
  4. Repeats up to 5 iterations until confident or max iterations reached

​
Citation Format

All facts include [10K-N] citation markers:
  • [10K-1] = First 10-K chunk retrieved
  • [10K-5] = Fifth 10-K chunk
  • Citations link to the exact section, table, or text passage with ticker, fiscal year, and section metadata

​
Coverage

  • Filing Type: 10-K only (10-Q and 8-K coming soon)
  • Years: 2018-2025 (varies by company)
  • Companies: S&P 500 and major public companies
  • Database: PostgreSQL with pgvector for text chunks, structured JSONB for tables

​
Technical Details

​
Architecture

User Question → Main Agent (semantic routing)
  ↓
[Routes to SEC Agent if 10-K data needed]
  ↓
SEC Agent:
  1. Plan Investigation (sub-questions + search plan)
  2. Parallel Multi-Query Retrieval (6 workers)
     - Table searches via LLM selection
     - Text searches via hybrid + reranking
  3. Generate Answer (with [10K-N] citations)
  4. Evaluate Quality (target 90%+)
  5. [If gaps] → Replan → Retrieve again (max 5 iterations)
  ↓
Return to Main Agent → Final response with [10K-N] citations

​
Search Strategy

For Financial Line Items (revenue, COGS, assets, etc.):
  • Type: table
  • LLM selects tables from income statements, balance sheets, cash flow statements
  • Full table content provided (never truncated)
For Narrative Data (business description, risks, workforce):
  • Type: text
  • Hybrid vector (70%) + keyword (30%) search
  • Cross-encoder reranking for relevance
  • Section filtering (e.g. Item 1 for business, Item 1A for risks)

​
Best Practices

1

Know When to Use 10-K vs Transcripts

  • 10-K: Annual data, audited financials, CEO compensation, full-year results
  • Transcripts: Quarterly commentary, management outlook, recent performance
2

Specify Fiscal Year

Use explicit fiscal years (“fiscal 2024” or “FY2024”) since companies define fiscal years differently.
3

Ask for Calculations

The agent handles derived metrics automatically:
  • “Calculate gross margin” → retrieves revenue + COGS, computes margin
  • “Debt-to-equity ratio” → retrieves debt + equity, computes ratio
4

Use Comparative Queries

“Compare X’s total assets in 2023 and 2024” → agent retrieves both years and presents side-by-side

​
Limitations

  • 10-K Only: No 10-Q (quarterly) or 8-K (current events) support yet
  • Annual Data: 10-K filings are annual, not quarterly (use earnings transcripts for quarterly data)
  • Fiscal Year Confusion: Companies define fiscal years differently (Apple FY2024 = Oct 2023 - Sep 2024)
  • No Intra-Year Updates: 10-K data is as of fiscal year-end, doesn’t reflect mid-year changes

​
Benchmark Performance

FinanceBench Evaluation (112 10-K questions):
  • Accuracy: 91%
  • Speed: ~10 seconds per question
  • Method: LLM-as-a-judge evaluation

Build docs developers (and LLMs) love

Get started for freeTalk to us