Advanced Financial Trend Analyst Agent

Dual-Tool Agentic Reasoning & Custom API Integration

This project demonstrates proficiency in building complex, self-hosted AI Agents capable of dynamic tool selection between quantitative database analysis (SQL) and qualitative compliance retrieval (RAG). This fully satisfies all application requirements regarding n8n (self-hosted), LLM-RAG, and custom API integration.

Project Summary (For Resume / Portfolio)

Feature	Technical Implementation
Custom Dual-Tool Agent	Developed a Gemini 2.5 Pro Agent that performs dynamic reasoning to select between a Custom SQL Executor (for data analysis/calculations) and a Custom Python RAG Retriever (for policy documents).
Advanced RAG & Tooling	Built a robust RAG stack using Ollama/PGVector with custom Python Code Node chunking for text pre-processing, ensuring superior RAG accuracy and bypassing standard node limitations.
API Integration	Implemented a Custom SQL Executor for database interaction and utilized the Telegram API for real-time reporting, demonstrating advanced API control over the workflow.

---

Technical Stack & Architecture

Component	Tool / Technology	Purpose
Workflow Orchestration	n8n (Self-Hosted)	Manages the end-to-end multi-step workflow.
LLM / Reasoning	Google Gemini 2.5 Pro	Core decision-making engine for tool selection and summarization.
Database & Vector Store	PostgreSQL with pgvector	Stores both transactional data (SQL) and policy embeddings (RAG).
Embeddings	Ollama (nomic-embed-text:latest)	Generates vectors for RAG document indexing.

---

RAG Ingestion Pipeline (Step 1: Data Preparation)

This workflow runs first to prepare the knowledge base, ensuring the Agent has data to retrieve:

1. File Read & Extract: Reads the policy file from a dedicated Docker volume (/rag_data) and converts the binary output to a usable text string.
2. Custom Chunking: Uses a Custom Python Code Node for text splitting (by double newline) to ensure logical context is preserved in each chunk.
3. Loading: Inserts chunks and Ollama-generated vectors into the financial_policy_rag table (addressing the column schema mismatch by renaming content to text).

---

Agentic Reasoning Workflow (Step 2: The Service)

This workflow operates continuously, using the data prepared above to service user requests.

Mechanism

The Agent uses a structured System Prompt to analyze the user's query and trigger one of two tools based on intent:

• Quantitative Query: $\rightarrow$ Calls the sql_query_executor tool, forcing the LLM to generate raw SQL (SELECT...).
• Policy Query: $\rightarrow$ Calls the custom_rag_retriever tool (Python), which performs a vector similarity search in the PGVector store.

Final Output

• The final synthesized summary is automatically delivered to the user's Telegram chat via the Telegram API (the external service integration).

---

Setup & Local Deployment

To run this project, follow these steps:

1. Clone Repository: git clone [repository URL]
2. Dependencies: Ensure Docker Desktop and Ollama (nomic-embed-text:latest) are running locally.
3. Start Services: Run docker compose up -d in the project root.
4. Configure Credentials: Add your Gemini API Key, Postgres, and Telegram Bot Token in the n8n Credentials UI.
5. Run Ingestion: Execute the RAG Ingestion: Financial Policy Load workflow once to populate the database.
6. Activate Agent: The Financial Analyst Agent workflow is now ready to respond via the Chat Trigger.

---

License

This project is protected under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International Public License (CC BY-NC-SA 4.0).

This license allows for sharing and adaptation only under the condition of attribution and strictly prohibits commercial use of the material, preserving the uniqueness of this solution.

The full legal text is available in the LICENSE file in this repository.