Introduction

As a treasury professional, you’re managing complex financial operations daily – cash forecasting, FX risk, investment strategies, and more. While AI assistants can improve efficiency, using public AI platforms risks exposing sensitive financial data. This guide demonstrates how to deploy a private GPT model with RAG (Retrieval Augmented Generation) capabilities specifically for treasury operations without coding knowledge or IT department involvement.

This guide uses AnythingLLM, which provides document management and RAG functionality, making it ideal for treasury professionals who need to work with large volumes of financial documents. Of course, you can try also LM Studio combined with Ollama, Mistral or DeepSeek, but RAG is quite difficult to enable if you are not very technical. You choose.

Why Private GPT with RAG Matters for Treasury

Corporate treasury deals with highly confidential information:

Cash position data across multiple entities
FX exposure details
Investment portfolio strategies
Banking relationship details
Credit facility terms

Using public AI platforms could potentially expose this data to third parties or allow it to be used for model training. A private deployment ensures your sensitive financial data remains entirely within your control.

You can shut down the internet connection and you became 100% private. Just kidding. Or not.

RAG capabilities are particularly important for treasury applications because they allow your AI to:

Reference your specific treasury policies
Analyze your actual cash management documents
Provide recommendations based on your company’s unique financial structure
Consider your existing banking relationships and terms

But what is RAG?

RAG = Retrieval-Augmented Generation. It’s a technique that allows an AI model (like ChatGPT, LLaMA, Mistral, etc.) to generate answers based on your documents, not just what it was trained on.

In a nutshell:

Without RAG	With RAG
The AI responds only based on its pretraining (up to a certain date)	The AI responds using information from your documents
It may guess or make mistakes when it lacks knowledge	It can quote accurately from PDFs, Excel files, Word docs, etc.
It doesn’t know what “treasury_policy.pdf” is	If you give it that file, it knows how to search it and answer accordingly

How does RAG work?

You upload your documents (PDF, Word, TXT, websites, etc.)
They are broken down into small text chunks
When you ask a question, the AI looks for the most relevant chunks
Then it generates an answer based on those chunks

Key Benefits

Fully local – RAG works even without an internet connection
Use your own files – confidential or internal documents
Fewer hallucinations – more accurate, grounded answers (but still has!)
More control – you decide what the AI can see and respond with

The Solution proposed: AnythingLLM with Open-Source Models

This combination provides a user-friendly approach that runs entirely on your computer without sending any data to external servers.

System Requirements

Before getting started, ensure your computer meets these minimum specifications:

Windows 10/11 or macOS 12+ (Windows recommended for treasury applications)
16GB RAM (32GB recommended for better performance)
30GB free storage space
For best performance: A modern CPU with 8+ cores

Step-by-Step Deployment Guide

1. Installing AnythingLLM (aprox 5-10 minutes)

Step 1: Download and install AnythingLLM

Visit AnythingLLM page
Click “Download for desktop” and choose your operating system.
Run the installer and follow the on-screen instructions

Step 2: Launch AnythingLLM

Start the application after installation completes

2. Setting Up Your LLM Model (20-30 minutes)

For treasury work, you’ll need a capable LLM. Llama 3 is recommended due to its strong performance in financial analysis:

Now, AnythingLLM offers several LLM provider options to power your private treasury GPT solution. For a treasury professional with limited technical expertise, here are my recommendations:

For a fully private, on-device solution:

AnythingLLM (powered by Ollama) is your best option from the list. This allows you to run models from Meta, Mistral and others directly on your device with zero setup, keeping all your treasury data completely private.

For the specific model to use with this option, I’d recommend:

Llama 3 8B Instruct if your computer has modest specifications (16GB RAM)
Llama 3 70B Instruct if you have a powerful computer (32GB+ RAM, good GPU)
Mistral Large as an alternative if you find Llama 3 performance unsatisfactory

If absolute performance is more important than complete privacy:

Anthropic (Claude) would provide excellent financial analysis capabilities, but your queries would be sent to Anthropic’s servers
Gemini (Google’s model) offers strong capabilities but also processes data on Google’s servers

For most treasury professionals, the AnythingLLM option with a locally-run Llama 3 model provides the best balance of privacy, performance, and ease of use. The model would stay completely on your device while still providing sophisticated financial analysis capabilities.

Step 3: For this demonstration, I’ve chosen Llama 3 as a model.

In AnythingLLM, in “LLM Preferences”, search for “llama3” in the search box
For most treasury users, select:
- “Meta-Llama-3-8B-Instruct” if your computer has 16GB RAM
- “Meta-Llama-3-70B-Instruct” if your computer has 32GB+ RAM and strong GPU
Click the right arrow on the screen, fill the fields, name your first workspace and wait for the process to complete (15-30 minutes depending on internet speed)

To add more models, you need to click Settings and then AI Providers here:

Step 4: When your model is downloaded, click here:

and set “LLM Temperature” to 0.2 for more precise financial recommendations. Lower temperature settings – like 0.2 – produce more deterministic, predictable outputs that are less likely to introduce creative variations or speculative content.

3. Creating Your Treasury Knowledge Base (45-60 minutes)

This critical step will provide your AI with your organization’s specific treasury knowledge.

Step 5: Prepare your treasury documents (at the end of this article, you can find some demo files you can download and use).

Export treasury policies as PDFs
Save relevant financial reports
Export cash position templates and recent reports
Download FX strategy documents
Include investment guidelines and current allocations
Include bank account structure documentation
Add recent treasury committee presentations
Include key banking relationship details

Step 6: Import documents into AnythingLLM

In AnythingLLM, click on “Workspaces” in the left sidebar
Choose your workspace, eg “Treasury GPT”
Inside your new workspace, click “Upload Files”
Select all relevant documents and add them (drag & drop)

Step 7: Write your instructions to your specialized treasury workspaces

Click on Settings button, go to Chat Settings, and add your Prompt
Create additional specialized threads where you discuss specific topics: Cash Management Advisor, FX Risk Specialist, etc. This allows for more focused expertise in each area.

Step 8: Second option: set up workspaces for each topic

You can create a workspace for each topic adding Prompt instructions for each one, documents, etc.

System Instructions – examples

Let’s pretend you create a workspace for each topic (one for Cash, one for FX, etc). See some prompt examples you could add:

Cash Management System Instructions:

You are a Treasury Cash Management Specialist with expertise in global cash pooling, forecasting, and liquidity management. 

Your primary responsibilities include:
1. Analyzing cash position reports to identify optimization opportunities
2. Reviewing cash forecasts for accuracy and suggesting improvements
3. Recommending cash concentration structures
4. Identifying idle cash and suggesting better utilization
5. Monitoring working capital metrics

Base your advice on the corporate treasury documents in the knowledge base and conservative treasury best practices. Always suggest practical, implementable solutions that minimize risk while optimizing returns.

When analyzing cash forecasts, look for:
- Historical accuracy patterns
- Seasonal variations
- Correlations between business units
- Anomalies that might indicate reporting errors
- Cash conversion cycle improvements

For cash pooling recommendations, always consider:
- Legal entity restrictions
- Regulatory limitations in each jurisdiction
- Tax implications of cross-border movements
- Bank fee structures
- Notional vs. physical pooling trade-offs

FX Risk Management System Instructions:

You are a Treasury FX Risk Specialist with expertise in currency risk identification, hedging strategies, and exposure management.

Your primary responsibilities include:
1. Analyzing FX exposure reports
2. Recommending appropriate hedging instruments (forwards, options, swaps)
3. Balancing risk mitigation with hedging costs
4. Suggesting improvements to FX risk policies
5. Helping quantify potential FX impacts on financial statements

Base your advice on the corporate treasury documents in the knowledge base and conservative treasury risk management principles. Always consider compliance with the company's risk policy limits and accounting implications of hedging strategies.

When analyzing FX exposures, consider:
- Translation vs. transaction exposures
- Natural hedging opportunities
- Correlation between currency pairs
- Historical volatility patterns
- Hedge accounting requirements

For hedging recommendations, always evaluate:
- Cost of hedging vs. potential risk
- Accounting treatment (cash flow hedge, fair value hedge)
- Impact on financial covenants
- Risk of over-hedging
- Appropriate tenor based on exposure certainty

Investment Portfolio System Instructions:

You are a Treasury Investment Specialist with expertise in short-term and medium-term investments, yield optimization, and liquidity management.

Your primary responsibilities include:
1. Analyzing investment portfolio composition
2. Recommending investment strategies within policy constraints
3. Optimizing yield while maintaining required liquidity
4. Balancing risk and return for different time horizons
5. Monitoring market conditions affecting investment decisions

Base your recommendations on the corporate investment policy documents in the knowledge base and conservative treasury investment principles. Always prioritize capital preservation and liquidity over yield enhancement.

When analyzing investment opportunities, consider:
- Current and projected interest rate environments
- Yield curve shape and implications
- Credit spread trends
- Liquidity needs by time horizon
- Counterparty risk concentrations

For portfolio optimization, always evaluate:
- Duration matching with liability profile
- Sector allocation within policy limits
- Credit quality distribution
- Instrument diversification
- Mark-to-market impact scenarios

Step 9: Configure RAG settings for optimal retrieval

In AnythingLLM, the RAG (Retrieval-Augmented Generation) mechanism is enabled by default—and it’s exactly what happens when you upload files and ask questions in a Workspace.

What does RAG mean in AnythingLLM?

Your files are converted into embeddings and stored in a vector database (see the “Vector Database” tab). When you ask a question, the AI searches for relevant snippets in those files → sends them along with your prompt to the LLM → and the LLM generates a more informed answer.

You can configure the RAG behavior through:

1. Vector Database Settings:

Search Preference: Leave it on Default, or choose another option if you want more precise results.
Max Context Snippets: Controls how many file snippets are sent to the LLM per query (recommended: 4–6).
Similarity Threshold: Defines how closely a snippet must match your question to be considered relevant:
- Low (≥ 0.25) = More snippets pass through (may lead to hallucinations).
- Medium/High (≥ 0.5 / 0.75) = Only highly relevant snippets are used.

2. Custom Prompt (under “Chat Settings”) – you can force the LLM to ignore anything outside the uploaded documents:

Only respond based on the uploaded documents. If there is no relevant information in the retrieved context, clearly say you cannot answer. Do not hallucinate or add knowledge beyond what is provided.

Want a “strict RAG” setup for a professional use case?

Here’s a recommended configuration:

Setting	Recommended Value
Max Context Snippets	5
Similarity Threshold	Medium (≥ 0.5)
Prompt	Only respond based on uploaded documents
LLM Temperature	0.2 (for accurate, factual responses)

Enhancing Your Treasury AI with Advanced Features

Step 10: Set up conversation memory

You can set up a form of conversation memory in AnythingLLM, but it’s not full long-term memory (like in ChatGPT for example). It’s more like short-term context recall based on a fixed number of past messages. To configure it,

In the Chat Settings tab, you’ll see

This is where you control how much of the recent conversation is remembered and passed to the LLM as context.

Default/recommended value: 20
Max recommended: ~45 (depending on model size and token limits)
Too high can cause response failures or truncation.

This is basically your temporary memory per thread — once you close the thread or exceed the token limit, earlier messages will no longer be included.

What it’s not:

It’s not persistent memory across threads.
It doesn’t “learn” from your previous conversations across sessions.
It doesn’t retain facts unless they’re reloaded via files or included in prompts.

Want longer memory?

You can simulate it by:

Uploading a transcript of previous convos as a file to your workspace.
Using tools like MemoryGPT (if integrated) or a custom embedding setup.
Increasing the Chat History value to hold more back-and-forth messages (within token limits).

Testing Your Treasury AI with Real-World Scenarios

Now it’s time to test your AI with sophisticated treasury scenarios to ensure it’s providing valuable insights.

For each specialized workspace, run 2-3 test scenarios
Upload relevant documents for each test
Evaluate responses for accuracy, relevance, and actionability

Real-World Treasury Application Scenarios

Scenario 1: Multi-Currency Cash Pooling Structure Optimization

Challenge: Your multinational corporation operates across 18 countries with 47 bank accounts in 12 currencies. You need to optimize your cash pooling structure to reduce idle balances while meeting local regulatory requirements and minimizing tax implications.

Test Procedure:

Upload to your Cash Management Advisor workspace:
- Current bank account structure diagram
- Last 6 months of daily cash positions by entity
- Local regulatory requirements by country
- Intercompany loan documentation
- Current banking fee schedules
Ask your AI: “Analyze our current global cash pooling structure and recommend a revised architecture that would:
1. Reduce idle cash balances by at least 15%
2. Maintain compliance with all local regulatory requirements
3. Minimize potential tax leakage from cross-border movements
4. Optimize banking fees across our relationship banks
5. Consider the impact of the new physical header account in Singapore
Include specific recommendations for:
- Header-account locations and currencies
- Physical vs. notional pooling structures by region
- Intercompany loan documentation requirements
- Recommended bank partner capabilities by region
- Implementation timeline and critical path items”

Expected Output: Your AI should provide a comprehensive analysis including:

Detailed current state assessment identifying inefficiencies
Recommended pooling hierarchy with specific account restructuring
Regulatory compliance considerations by jurisdiction
Tax-efficient movement strategies
Quantified potential interest savings and fee reductions
Implementation plan with key milestones and dependencies
Required legal entity and documentation changes

Scenario 2: Dynamic FX Hedging Strategy for Volatile Revenue Streams

Challenge: Your organization has significant but irregular FX exposures from project-based contracts in emerging markets with high currency volatility. You need a sophisticated hedging approach that balances protection against downside risk while preserving opportunity for favorable movements.

Test Procedure:

Upload to your FX Risk Specialist workspace:
- Current FX exposure report by currency pair
- Historical project contract payment schedules (last 24 months)
- Current hedging positions and instruments
- FX volatility data by currency pair
- FX policy risk limits
- Hedge accounting documentation requirements
Ask your AI: “Develop a dynamic layered hedging strategy for our project-based exposures in BRL, INR, and ZAR that:
1. Protects against downside beyond our 15% risk tolerance threshold
2. Maintains hedging ratio between 50-75% of confirmed exposures
3. Provides flexibility to adjust as project timelines shift
4. Balances premium costs against protection value
5. Complies with our hedge accounting requirements
Include recommendations for:
- Specific instrument mix (forwards, collars, options) by tenor
- Trigger points for hedge ratio adjustments
- Maximum option premium budget allocation
- Correlation analysis between our exposure currencies
- Stress testing scenarios and potential P&L impacts
- Economic vs. accounting treatment considerations”

Expected Output: Your AI should provide a sophisticated analysis including:

Currency pair-specific hedging strategies
Layered approach with appropriate instrument mix
Trigger-based adjustment methodology
Premium budget optimization analysis
Accounting treatment considerations
Stress test scenarios showing potential outcomes
Correlation-based portfolio effects
Specific recommended trades with notional values and tenors

Maintaining Your Treasury AI System

Establish regular maintenance procedures like:

Schedule bi-weekly document updates:
- Remove outdated treasury documents
- Add new policies or guidelines as they’re approved
- Update performance data and exposure reports
Check monthly for AI model updates:
- In AnythingLLM settings, check “Available Updates”
- Update to newer Llama models as they become available
Document effective query patterns:
- Keep a shared document of effective prompts and questions
- Note which question formats yield the most actionable insights

Performance Optimization Tips

For complex treasury analyses: Break questions into logical components
For document-heavy work: Consider creating separate workspaces for different document categories
For slow responses: Reduce the context window or number of retrieved chunks
For better answers: Be specific in your questions and reference relevant policy sections
For team adoption: Create a “prompt library” of effective questions for different treasury scenarios

Conclusion

You’ve now deployed a private, secure Treasury GPT with powerful RAG capabilities using AnythingLLM and a Llama 3 model. This approach keeps your sensitive financial data completely under your control while providing sophisticated AI assistance for complex treasury operations.

The RAG capabilities are particularly valuable for treasury work as they allow your AI to reference your specific policies, analyze your actual financial data, and provide recommendations tailored to your organization’s unique treasury requirements.

Your Treasury AI system will become increasingly valuable as you:

Add more relevant treasury documents to your knowledge base
Refine your system instructions with domain-specific expertise
Build a library of effective question patterns
Update with the latest AI model capabilities

By keeping all processing local and data private, you can confidently use AI to transform your treasury operations without compromising security or confidentiality.

Disclaimer

This article provides a technical framework and demonstrates how to leverage existing tools to streamline treasury processes. The implementation described here assumes:

Appropriate governance frameworks are already in place within your organization.
Human oversight remains essential at critical approval points as detailed in the workflow.
Organization-specific controls must be integrated based on your company’s policies and risk tolerance.

This guide focuses exclusively on the technical implementation aspects rather than governance, compliance, or accountability frameworks, which will vary by organization/ country/ region. Always consult with your compliance, security, finance, and legal teams to ensure the solution meets your organization’s specific requirements and standards before implementation.

Any automation solution should enhance—not replace—human judgment in financial processes. I don’t suggest full automation of processes, but rather the streamlining of workflows while maintaining appropriate controls.

Deploying Your Private Treasury GPT with RAG: A Complete Guide for Non-Technical Users