Sachin Rekhi Outlines 5 Key Strategies for AI Workflow Context

In a recent LinkedIn post, Sachin Rekhi delves into a critical challenge faced by many in the burgeoning field of artificial intelligence: establishing an effective context strategy for AI workflows. Rekhi emphasizes that the most difficult aspect of building these systems today is not the AI model itself, but rather determining how to supply it with the necessary data to perform its tasks.

He outlines five distinct strategies that users can employ to ensure their AI workflows have access to the right information. According to Rekhi, the choice of strategy often depends on the specific requirements of the workflow, balancing factors like speed, reliability, and ease of implementation.

Understanding the Core Challenge: Context is King

Rekhi frames the central problem as a ‘context strategy’ – essentially, the method by which an AI gains access to the data it needs. He elaborates on this, stating:

“The hardest part of building an AI workflow today is deciding your context strategy, which is how are you going to get the data you need for the task?”

This fundamental issue, as highlighted by Rekhi, underpins the success or failure of many AI applications. Without a robust method for providing relevant data, even the most sophisticated AI models will struggle to deliver accurate or useful results.

Five Strategies for AI Data Acquisition

To address this challenge, Rekhi details five practical approaches, each with its own advantages and disadvantages:

1. Local Files

Rekhi identifies using local files as the ‘fastest and most reliable’ method. This involves workflows directly accessing data stored on the user’s device. As an example, he mentions utilizing downloaded markdown meeting notes for drafting agendas, which allows the AI to quickly reference past discussions.

2. CLI Tools

Command-line interface (CLI) tools are presented as another highly efficient option. Rekhi notes that AI tools are adept at interacting with CLIs, which are prevalent for a vast array of functions. He explains:

“AI tools are incredibly good at running command-line tools, which are programs that run in the Terminal. CLIs exist for pretty much everything, they are very fast to run, and quite reliable.”

An example provided is using the Whisper CLI for transcribing video interviews, showcasing the power and speed of this method.

3. MCP Servers

For accessing remote content, Rekhi suggests MCP (Multi-Channel Processing) servers. These facilitate connections to cloud-based platforms like Google Docs, Notion, and Slack. Rekhi uses a Slack MCP server in his own workflow to stay updated on channel discussions. However, he expresses a growing preference for CLI tools when available, citing speed and reliability concerns with MCP servers.

4. APIs

When neither a CLI nor an MCP solution exists, Rekhi turns to Application Programming Interfaces (APIs). This involves instructing the AI to write code that interacts with an API to retrieve data. While this method offers broad data access, Rekhi cautions that it requires more setup and can be prone to bugs. He reserves this approach for situations where other options are not viable, such as using the Gamma API for generating presentations from NPS analysis data.

5. Browser Agent

The final strategy, and one Rekhi considers a last resort, is the browser agent. This allows AI tools to navigate websites, click links, and extract information. Rekhi describes it as:

“the slowest and least reliable method. So I only turn to it when there are literally no other options.”

He cites using this technique to scrape competitor pricing pages when no other data access method was feasible.

Conclusion: A Strategic Approach to AI Data

Sachin Rekhi concludes his post by encouraging users to consider these five strategies when developing their AI workflows. By understanding the trade-offs between speed, reliability, and implementation effort, individuals can build more effective and efficient AI systems tailored to their specific needs.