Rethinking Research: Private GPT for Investment Analysis

This allows chatbot-style equipment analysts to query complex research materials in plain language without highlighting sensitive data in the cloud.

Case for “Private GPT”

For professionals working in by-side investment research-the use of an equity, fixed income, or multi-asset strategies-Chatgpt and similar tools are a major concern: privacy. Uploading documents offered by research reports, investment memmo, or drafts for cloud-based AI tools is usually not an option.

This is the place where “private GPT” comes: A structure is made perfectly on open source components, which is running locally on its machine. Application programming interface (API) has no dependence on the key, there is no need for internet connection, and there is no risk of data leakage.

This toolkit takes advantage:

• Python scripts For ingestion and embeding of text documents
• AlaamaAn open-source platform for hosting local LLM on computer
• streamlined To manufacture user friendly interface
• Mistral, DeepsekAnd Other open-source models To answer questions in natural language

For this example, the underlying Python Code is publicly placed here in the Github Repository. Additional guidance on step-by-step implementation of technical aspects in this project has been provided in this assistant document.

Do research like a chatbot without cloud

The first step in this implementation is starting a pythan-based virtual atmosphere on an individual computer. It helps to maintain a unique version of package and utilities that feed in this application alone. Consequently, the settings and configurations of packages used in pythons for other applications and programs are undivided. Once installed, a script reads and embed the investment documents using an embeding model. These embedding allow LLM to understand the content of the document at a granular level, aimed at occupying the semantic meaning.

Because the model is hosted via Olama on a local machine, the documents remain safe and do not leave the analyst’s computer. This is particularly important when proprietary research, non-public financial financial transactions or internal investment notes are dealt with notes.

A practical display: analysis of investment documents

The prototype focuses on long -term investment documents such as earnings call tape, analyst report, and offering statement. Once the TXT document is loaded into a specified folder of the individual computer, the model processes it and is ready to interact. This implementation supports various types of document types ranging from Microsoft Word (.Docx), website page (.HTML) to PowerPoint Presentation (.PTX). The analyst may start querying the document through a model selected in a simple chatboat-style interface provided in a local web browser.

Using a web browser-based interface operated by streamlight, analyzer may start querying the document through the chosen model. Even though it launches a web-bruser, the application does not interact with the Internet. Browser-based rendering is used to display a convenient user interface in this example. This can be modified in command-line interfaces or other downstream manifestations. For example, after completing AAPL’s earnings call transcript, one can simply ask:

“What do Tim Cook do in Aapl?”

Within seconds, LLM transcripts and return materials from returns:

“… Timothy Donald Cook is the Chief Executive Officer (CEO) of Apple Inc. …”

This result is cross-satisfied within the tool, which also shows from which pages the information was drawn. Using the mouse click, the user can expand the “source” item listed below each response in the browser-based interface. Various sources feeding in that answer are ranked based on relevance/importance. The program can be modified to list a different number of sources references. This feature enhances transparency and confidence in the output of the model.

Model switching and configuration for extended performance

A standout feature has the ability to switch between different LLMs with one click. The performance displays the capacity of the cycle between open-sources LLMs such as Mistral, Mixtral, Lama and Dipsek. This suggests that different models can be plugged into the same architecture to compare performance or improve results. Olama is an open-source software package that can be installed locally and facilitates this flexibility. As more open-sources models are available (or the current updates), Olma is able to download/update accordingly.

This flexibility is important. This allows analysts to test which models correspond to the nuances of a particular task in the hand, that is, without the need for access to the legal language, financial revelations, or research summary, all payment APIs or enterprise-wide licenses.

There are other dimensions of the model that can be modified to target better performance for a given task/purpose. These configurations are usually controlled by a standalone file, usually named as “config.py” in this project. For example, the equality range between the text of the text in a document can be modified to identify very close matches using high value (more than 0.9). This helps in reducing noise, but if the threshold for a chosen reference is too tight, then can miss the results related to semantics.

Similarly, the length of minimal chunk can be used to identify and weed a very small piece of text that are unexpected or misleading. Important ideas also arise from the shape of the chunk and overlap options between the fragmentation of the text. Together, they determine how the document is divided into pieces for analysis. Larger size sizes allow for more reference per answer, but can also dilute the attention of the subject in the final response. The amount of overlap ensures smooth continuity between subsequent fragmentation. This ensures that the model can explain information that extend in many parts of the document.

Finally, the user will also have to determine how many parts of the text should be focused for the final answer among the recovered top objects for the query. It leads to balance between speed and relevance. Using too much target fragmentation to each query response can slow the device to slow down and feed in potential distractions. However, using very low goal fragmentation may lead to the risk of missing significant references that cannot always be written/discussed in geographical proximity within the document. In combination with various models served through Olama, the user can configure the ideal settings of these configuration parameters to suit their function.

Scaling for research teams

While the performance originated in the Equity Research Space, the implications are broad. Fixed income analysts can offer statements and contractual documents related to Treasury, Corporate or Municipal Bond. Macro researchers can swallow Federal Reserve Speech or Economic Outlook documents from central banks and third -party researchers. Portfolio teams may pre-load the memo or internal report of the investment committee. Purchase-side analysts can use large versions of research exclusively. For example, hedge funds, martial vesses, process 30 petbytes each day equivalent to about 400 billion emails.

Accordingly, the overall process in this structure is scalable:

• Add more documents to the folder
• Prepare the embeding script swallowing these documents again
• Start interacting/query

All these stages can be executed in a safe, internal environment, which costs nothing to operate beyond local computing resources.

By putting AI in the hands of analysts – safely

The rise of generic AI does not require to surrender data control. By configuring the open-source LLM for private, offline use, analysts can create an in-house application such as chatbott here which are only capable than some professional options-and infinitely more secure.

This “private GPT” concept empowers investment professionals:

• Use AI for document analysis without highlighting sensitive data
• Reduce dependence on third-party equipment
• Tailor to the system for specific research workflows

The full codebase for this application is available on github and can be extended or sewn for use in any institutional investment settings. There are several points of flexibility in this architecture that enable the end-user to apply their choice for a specific use case. The underlying features regarding the check -up of the source of reactions help to detect the accuracy of this tool, so that to avoid the general loss of hallucinations between LLMS. This repository is to work as a guide and the initial point for the construction of downstream, local application that are ‘fine’ for enterprise-wide or personal needs.

Generative AI does not need to compromise on privacy and data security. When used carefully, it can increase the capabilities of professionals and help them analyze faster and better information. Such devices put AI directly into the hands of analysts-no third-party license, no data compromise, and no business between insight and security.