Automate SDLC with Gemini API and Vector Search

What this is, in plain English

On most software teams, someone turns a vague idea (“let users reset their password”) into detailed requirements, a technical plan, and a list of tasks. This workflow automates that, and crucially, it grounds the AI’s output in your actual codebase so the tasks fit the code you already have, instead of generic advice.

The trick that makes it work is “vector search”. Your code is converted into lists of numbers (“embeddings”) that capture meaning, so the system can find the parts of your codebase most related to a new feature and feed them to the AI as context. That technique (find relevant material, then let the AI use it) is called RAG, Retrieval-Augmented Generation.

Be honest about the level: this is an advanced, enterprise-flavored setup. It wires together several paid Microsoft Azure services, an AI model, and an automation tool, and the exact build is not a copy-paste recipe. This page explains what it does, where it is worth it, and the realistic shape of building it.

What you can use it for

• Consistent specs from one-line ideas. Turn short user stories into full, uniform requirements and technical specs every time.
• Tasks that fit your real code. Generate to-do items that reference the modules and patterns you actually use, not generic ones.
• Faster, steadier planning. Keep a small team’s backlog detailed and current without a person writing every ticket.
• Onboarding and documentation. Produce living docs grounded in the codebase, useful for new hires.
• The reusable idea (RAG over your own data) applies well beyond code: ground an AI in your contracts, policies, or product manuals.

Tools you need

• Gemini API (freemium): the AI that writes the requirements, specs, and tasks. Has a free tier.
• Azure OpenAI (paid): used to turn your code into embeddings (the number-lists that capture meaning).
• Azure AI Search (paid): stores those embeddings and finds the most relevant code for a given request.
• Power Automate (freemium): the automation tool that runs the steps in order, calling each service.
• Azure DevOps (freemium): your project tracker, where the generated tasks land as work items.

How it actually works

This is a pipeline you assemble, not a quick recipe. The realistic shape:

1. Stand up Azure AI Search. Create the search service that will hold your codebase’s embeddings.

2. Turn your code into embeddings. Write a script that reads your code files, sends them to an Azure OpenAI embedding model, and gets back the number-lists.

3. Load the embeddings into Azure AI Search. Now your codebase is searchable by meaning, not just keywords.

4. Build the orchestration in Power Automate. Create a flow that will run the whole sequence whenever you feed it a user story.

5. Generate the documents with Gemini. The flow calls the Gemini API with a prompt for each piece (requirements, tech spec, and so on). A starting prompt:

   Given the following user story, generate a detailed set of requirements:
   [Paste the user story here]

6. Add RAG for context. Before each Gemini call, search Azure AI Search for the code most related to the request and paste those snippets into the prompt, so the output fits your actual code.

7. Create the tasks in Azure DevOps. The flow calls the Azure DevOps API (using a Personal Access Token, a scoped password for automation) to create the work items from the generated specs.

Words you’ll see, explained

• SDLC: software development lifecycle, the whole journey from idea to shipped feature.
• User story: a short, plain description of something a user wants (“as a user, I can reset my password”).
• Embedding: a list of numbers that represents the meaning of a piece of text or code, so a computer can compare them.
• Vector search: finding items whose embeddings are closest in meaning to your query.
• RAG (Retrieval-Augmented Generation): fetch the most relevant material first, then let the AI answer using it.
• Orchestrate: run a series of steps automatically in the right order (Power Automate’s job here).
• Personal Access Token (PAT): a scoped password that lets a script act in a service like Azure DevOps.

Original source

Based on a Hacker News post by Antony Brahin describing how they automated parts of their software development lifecycle by running vector search over their codebase to give the AI real context.

Notes & variations

• Do you even need this? For a solo project, prompting an AI with your code pasted in is far simpler. This level of machinery pays off for a team that produces many specs and wants them consistent and code-aware.
• Free-tier alternatives: open-source vector databases (like ChromaDB or Weaviate) and self-hosted embedding models can replace the paid Azure pieces, at the cost of much more setup.
• Common pitfall: vague prompts. Tell Gemini exactly what format and detail you want, or you get generic output that ignores your context.
• Tip for better results: tune how many code snippets you retrieve and how similar they must be. Too few starves the AI of context; too many drowns the real signal.