Research Agent (Apple 10-K Analysis Assistant)

English | [简体中文](./README.md)

An AI research assistant built with LangGraph for analyzing Apple Inc. SEC 10-K filings. It uses hybrid retrieval (keyword search + semantic reranking) with Milvus vector store for high-precision financial Q&A.

Agent and Retrieval Architecture Design

The project leverages LangGraph to design a multi-step retrieval and analysis pipeline specifically tailored for complex long texts like financial reports. It primarily involves three core components:

1. Data Pre-processing & Indexing

• Text Chunking: Structured Apple 10-K financial report JSON data is parsed and sliced using RecursiveCharacterTextSplitter (default 1500 chars limit, 150 chars overlap) to ensure coherent contextual semantics.
• Embedded Vectorization: The application utilizes Qwen3-Embedding-8B (via an OpenAI-compatible API) to bundle text chunks and calculate vector embeddings.
• Storage & Metadata: Vectors are stored in the Milvus vector database (handling Lite or Standalone environments). Comprehensive metadata attributes (such as symbol, fiscal_year, section_title, and form_type) are preserved alongside the vectors, establishing a foundation for precise structured condition filtering.

2. Text Retrieval & Information Augmentation (RAG Pipeline)

• Multi-dimensional Concurrent Retrieval: Based on the user's questions, inquiries are divided into different evaluation perspectives and keyword groups. Concurrent retrieval is executed across these keyword groups utilizing a ThreadPoolExecutor. This parallel design significantly enhances the recall efficiency of complex financial report contexts.
• Content Deduplication & Cleansing: The retrieval results from various threads are aggregated and strictly deduplicated via their page_content to prevent redundant information from interfering with large language model inference.
• Semantic Reranking: The recalled preliminary materials are distributed to the Qwen3-Reranker-8B model based on distinct evaluation perspectives (Perspectives) for deep semantic reranking and scoring. The high-matching structured Markdown content is then filtered out to complete the multi-dimensional information augmentation.

3. Answer Generation & Brief Analysis Reports

• Core Inference and Output: After obtaining highly condensed and cleansed background context from the reranking mechanism, the Agent's decision-making brain (Graph) integrates with powerful conversational LLMs (e.g., MiniMax-M2.7). Based on the user's questions and the financial report snippets, it conducts comprehensive analysis and tabular summaries. Ultimately, it can not only handle common Q&A but also automatically draft logically rigorous stage-by-stage brief financial analysis reports.

Architecture

• Agent Core: LangGraph-based agent (agent.graph:graph) with multi-step reasoning.
• RAG Pipeline:
  • Vector Store: Milvus (Standalone or Lite) with hybrid search.
  • Models: OpenAI-compatible APIs for Embedding and Reranking (e.g., Qwen3).
• Backend: FastAPI upload service (make api).
• Frontend: Next.js interactive UI (make ui).

Quick Start (Production/Full-Stack)

1. Environment Setup

Configure your API keys in the .env file:

cp .env.example .env

2. Build & Launch Infrastructure

Build the LangGraph server image before starting Docker:

make sync        # Install dependencies
make build       # Build production image via langgraph cli

Start the production stack (Milvus, etc.):

make up-prod

3. Data Ingestion

Once the stack is healthy, ingest the sample Apple 10-K data:

make ingest-prod # Process and upload aapl_10k.json to Milvus

4. Run & Interact

Start the UI and API servers:

• Frontend:

  make ui-install
  make ui          # Starts at :3000

• Backend/Agent:

  make api         # Starts at :8080
  make run         # Starts LangGraph dev server at :2024

Common Commands

Command	Description
make up-prod	Start production Docker stack (Milvus standalone)
make ingest-prod	Ingest Apple 10-K JSON into production Milvus
make dev	Start local dev flow (LangGraph + Frontend)

---

See CLAUDE.md for more technical details.