Andrii Bidochko

• Updated: July 24, 2026
• 5 min read

Augmenting Fundamental Analysis with Large Language Models: A RAG-Based System for Generating Investor Briefs

Direct Answer

The paper introduces a Retrieval‑Augmented Generation (RAG) pipeline that combines large language models with SEC‑EDGAR filings and macro‑economic data to automatically produce concise, investment‑grade briefs.

This matters because it demonstrates how generative AI can turn massive, unstructured financial disclosures into actionable insights, dramatically lowering the barrier for individual investors and finance teams to perform high‑quality fundamental analysis.

Background: Why This Problem Is Hard

Fundamental analysis traditionally relies on two labor‑intensive steps:

• Data acquisition: Extracting relevant numbers from thousands of pages of SEC filings, earnings calls, and macro‑economic releases.
• Insight synthesis: Translating raw figures into a narrative that highlights risks, growth drivers, and valuation implications.

Even seasoned analysts spend hours sifting through PDFs, spreadsheets, and news feeds. Existing automation tools either:

• Offer rule‑based extraction that breaks when document formats change, or
• Rely on generic language models that hallucinate numbers or misinterpret legal language.

Consequently, most retail investors lack timely, reliable briefings, while professional teams must allocate costly human resources to keep up with filing cycles. The problem is amplified by the sheer volume of quarterly reports—over 10,000 filings per quarter in the U.S. alone—making scalability a critical bottleneck.

What the Researchers Propose

The authors present a modular RAG framework that stitches together three core components:

1. Document Retriever: A vector‑based search engine that indexes SEC‑EDGAR filings, macro‑economic releases, and supplemental analyst commentary.
2. Structured Extractor: A fine‑tuned LLM (GPT‑4o) that parses retrieved documents, extracts key financial metrics (e.g., revenue, EBITDA, cash flow), and normalizes them into a tabular schema.
3. Brief Generator: A second LLM that consumes the structured data, applies a prompt template emphasizing investment relevance, and produces a 300‑word narrative brief.

Crucially, the pipeline is “retrieval‑augmented”: the generator never operates on raw text alone; it always receives a curated, factual grounding set, which dramatically reduces hallucination risk.

How It Works in Practice

The workflow can be visualized as a four‑stage loop:

1. Ingestion: New SEC filings are automatically downloaded via the EDGAR API and stored in a vector database.
2. Query Formation: An analyst or automated scheduler issues a request for a specific ticker and reporting period.
3. Retrieval & Extraction: The retriever pulls the top‑k most relevant sections; the extractor then converts them into a JSON payload of financial facts.
4. Generation: The brief generator receives the JSON payload plus a market‑context prompt (e.g., “Compare Q2 2024 results to the prior year and to sector averages”) and outputs a concise briefing.

What sets this approach apart is the strict separation between factual grounding (retrieval + extraction) and narrative creation (generation). By feeding the generator only verified numbers, the system sidesteps the most common failure mode of LLM‑only pipelines—fabricating data.

Evaluation & Results

The authors evaluated the pipeline on two fronts:

Quantitative Accuracy

• Extraction precision of key metrics reached 94 % against a manually annotated benchmark of 500 filing excerpts.
• Generated briefs contained zero numeric hallucinations in a blind test of 200 samples.

Investor Feedback

A user study with 120 individual investors compared the AI‑generated briefs to traditional analyst reports. Participants rated the AI briefs as:

• “More concise” (78 % of respondents)
• “Easier to understand” (71 %)
• “Equally trustworthy” (62 %)

Overall satisfaction scores were 0.6 points higher on a 5‑point Likert scale, indicating that the RAG pipeline delivers both speed and perceived quality.

Why This Matters for AI Systems and Agents

For practitioners building AI‑driven finance products, the paper offers a blueprint for integrating LLMs into regulated data pipelines without sacrificing compliance:

• Agent‑centric design: The modular components map cleanly onto autonomous agents—one for retrieval, one for extraction, one for synthesis—enabling orchestration via platforms like the Workflow automation studio.
• Scalable knowledge bases: Vector stores such as Chroma DB integration can handle millions of filing vectors, supporting enterprise‑level workloads.
• Compliance‑first architecture: By keeping raw source documents immutable and exposing only verified extracts, firms can audit the provenance of every figure used in a brief.
• Rapid prototyping: The same RAG stack can be repurposed for earnings call summarization, ESG report generation, or macro‑trend newsletters, extending the value of a single investment in the pipeline.

What Comes Next

While the results are promising, several limitations remain:

• Coverage breadth: The current study focuses on U.S. public companies; extending to private filings, international standards (e.g., IFRS) will require multilingual retrievers.
• Temporal reasoning: The generator does not yet perform multi‑period trend analysis beyond the immediate filing window.
• Regulatory scrutiny: Automated brief generation may trigger “investment advice” regulations in certain jurisdictions, necessitating built‑in disclaimer layers.

Future research directions include:

• Integrating real‑time market data feeds to contextualize financial metrics with price movements.
• Embedding domain‑specific ontologies to improve extraction of non‑standard line items.
• Deploying reinforcement‑learning‑from‑human‑feedback (RLHF) loops where professional analysts grade generated briefs, further aligning the model with industry standards.

Potential applications span from AI marketing agents that automatically craft investor newsletters, to Openclaw (Clawdbot, MoltBot) style bots that answer investor queries in natural language.

Conclusion

The Retrieval‑Augmented Generation pipeline presented in the paper demonstrates a practical path for turning dense regulatory filings into succinct, trustworthy investment briefs. By anchoring large language models to a verified retrieval layer, the approach mitigates hallucination, scales across thousands of companies, and delivers a user experience that rivals traditional analyst reports. For AI product teams, the modular design aligns with modern agent‑orchestration frameworks, opening the door to a new generation of AI‑enhanced financial tools.

As the finance industry continues to digitize, the ability to automatically synthesize high‑quality fundamental analysis will become a competitive differentiator. Organizations that adopt RAG‑based pipelines early—while addressing compliance and multilingual challenges—stand to empower both retail investors and professional analysts with faster, data‑driven insights.

{{IMAGE}}

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.