Skip to content
Archisys

Documentation · Archisys Technologies

SEC Research

SEC Research is an evidence-first SEC research engine for AI assistants and agents. It retrieves official SEC filings and financial data, identifies relevant filing sections, extracts and ranks evidence, and returns source-attributed results for downstream AI reasoning.

The current research engine uses deterministic retrieval and evidence-ranking. It does not use an LLM inside the research engine. The AI client or agent remains responsible for reasoning and synthesis.

This is not a trading agent, an autonomous investment advisor, a general-purpose LLM, or simply an SEC EDGAR API wrapper. It is evidence-first SEC research infrastructure: a deterministic MCP service that returns source-attributed evidence for downstream AI reasoning.

1. Overview

SEC Research is an Archisys Technologies product and a hosted MCP service for AI assistants and agents. It began as a retrieval and normalization layer over public SEC EDGAR data and has evolved into an evidence-oriented research engine: it selects filings, detects sections, extracts evidence, ranks that evidence deterministically, and returns source-attributed results.

Conversational interpretation happens in the client. SEC Research does not invent missing values, execute trades, manage portfolios, or give investment recommendations.

2. Architecture

The production path is a deterministic evidence pipeline exposed through MCP. Traffic is read-only against public SEC EDGAR sources. Semantic models and rerankers are not part of the production architecture.

SEC EDGAR

Official public filings, facts, and company identifiers

SEC Research MCP

Hosted Streamable HTTP research interface

Filing Selection

Identify the filing that matches the research request

Document Parsing / iXBRL Cleaning

Prepare filing text for section-aware retrieval

Section Detection

Locate relevant filing sections where supported

Evidence Chunking

Split cleaned sections into retrievable evidence units

Query Intent / Subject-Aware Retrieval

Target evidence to the research subject

Deterministic Evidence Ranking

Score and order evidence without an LLM in the engine

Source-Attributed Evidence

Return snippets and facts with filing provenance

AI Client / Agent

Downstream assistant or agent that consumes evidence

Reasoning & Synthesis

Performed by the AI client, not inside the MCP

3. Evidence engine

SEC Research is designed as an evidence layer rather than a raw EDGAR proxy. The research engine performs filing selection, document cleaning, section detection, evidence chunking, query-intent targeting, subject-aware retrieval, deterministic relevance scoring, and provenance tracking before evidence is returned to the downstream AI system.

The objective is not to generate an answer inside the MCP. The objective is to return trustworthy evidence that an AI system can reason over.

4. Evidence Before Reasoning

  • Retrieve the correct filing.
  • Identify the relevant filing section.
  • Find evidence relevant to the research subject.
  • Reject weak or irrelevant matches when they fall below the relevance threshold.
  • Preserve filing and source provenance.
  • Let the downstream AI system perform reasoning and synthesis.

The engine prefers returning no evidence over confidently returning unrelated evidence.

5. MCP endpoint

Live hosted MCP (Streamable HTTP):

https://sec-research.archisystech.com/mcp

Health check:

https://sec-research.archisystech.com/health
→ {"status":"ok"}

Muse connection

6. Available tools

These are MCP tools on the research engine. They are not AI agents.

search_company
Search SEC company identity and ticker/CIK information.
search_filings
Find relevant SEC filings and filing metadata.
get_filing
Retrieve filing content, with section-aware retrieval where supported.
get_company_facts
Retrieve normalized XBRL/company facts with deterministic concept resolution.
financial_history
Retrieve historical financial metrics from SEC data.
compare_companies
Compare financial metrics across companies with computed deltas. Does not rank companies.
filing_research
Perform evidence-oriented research across filing sections and return ranked, source-attributed snippets. There is no LLM inside the research engine.

7. Controlled Evaluation

The research engine is evaluated against a controlled 25-question benchmark covering filing retrieval, section targeting, financial facts, and filing-research questions.

Deterministic development has demonstrated major improvements in:

  • section accuracy
  • subject-aware evidence retrieval
  • rejection of irrelevant and nonsense queries
  • evidence ranking for common research questions

This evaluation is used to improve the engine. It is not a claim of complete accuracy.

8. Current vs research direction

Current

  • Deterministic SEC retrieval
  • Section-aware evidence extraction
  • Subject-aware ranking
  • Provenance
  • MCP interface
  • No LLM inside the research engine

Research direction

  • Semantic relevance reranking for cases where relevant evidence exists but is expressed through paraphrase or more complex semantic relationships

Not a current production feature. SEC Research is not a specialized SEC model.

9. Example prompts

  • Find Apple's latest 10-K.
  • What was Apple's revenue over the last five years?
  • Compare Nvidia and AMD's revenue.
  • Find Microsoft's latest 8-K filings.
  • What risks did Nvidia disclose in its latest 10-K?

10. SEC data sources

SEC Research uses official public SEC sources only:

11. Rate limiting

Access to SEC endpoints is not unlimited. The service uses a centralized limiter, retries with backoff, and honors Retry-After. A SEC User-Agent is required for requests to SEC systems.

12. Access and authentication

  • No authentication
  • No user account required
  • No billing
  • Read-only public SEC EDGAR data

Questions: support@archisystech.com

13. Limitations

  • Does not invent missing values
  • Annual and quarterly facts are not mixed
  • Source URLs are attached when available
  • compare_companies does not rank companies
  • filing_research returns ranked snippets; it does not run an LLM in the research engine
  • Does not execute trades, manage portfolios, or give investment recommendations
  • Public data may be incomplete, delayed, or contain errors originating at the source

14. Security

SEC Research is a read-only MCP service over public SEC EDGAR data. It does not collect account credentials. Use is subject to the Terms of Service and Privacy Policy.

Problems this page addresses

Browse all problems and pages