How do I use RAG in my daily coding and vibe coding?

You already use RAG every time your coding agent searches your code before it writes. To get more from it, give the agent three more places to look: today’s library docs, one short file with your team’s rules, and your tickets and wiki. Then check which page it used.

RAG (retrieval-augmented generation) in one line: the AI looks things up first, then answers from what it found. If that is new, my one-minute video What is RAG? shows it with a shelf of files. This article is about the shelf a developer has.

Your coding agent already searches your code

When I ask an agent to fix a booking bug, the first thing it does is not write. It runs searches: file names, the error text, the function that throws it. Then it opens three or four files and reads them. That is retrieval. The answer comes from your code.

There are two ways a tool can do this, and the tools have been moving between them.

An index. The tool cuts your code into chunks and stores each one as a vector, a list of numbers for its meaning (my video What is a vector database? shows how). A question finds code that means the same thing, even with different words. GitHub Copilot still works this way: its #codebase tool “is always semantic”.

Agent search. The agent uses grep and file reads, the same tools you use, and decides what to search next. Boris Cherny, who created Claude Code, wrote in February 2026: “Early versions of Claude Code used RAG + a local vector db, but we found pretty quickly that agentic search generally works better” (Boris Cherny on X). Codex is told to prefer rg (ripgrep, a fast grep). In July 2026 Cursor stopped storing embeddings of your code for search and moved to a local grep index. In November 2025 Cursor had published a 12.5% accuracy gain from semantic search. Eight months later, Cursor’s team wrote on its forum that models had become so good at grep that the index was no longer helping much.

The research is not settled. A May 2026 paper that tested Claude Code, Codex and Gemini CLI on a memory benchmark found grep “generally” more accurate than vector search. An August 2026 paper on questions about real repositories found the opposite: semantic search 65.2% right, deep agent search 46.2%. My rule of thumb:

You do not need to pick. Use the tool you have, and when the answer is wrong, first ask: did it look in the right file?

Give the AI today’s library docs, not its memory

A model knows a library as it was on its training cutoff day. Claude Sonnet 4.6 stops in August 2025 (Anthropic). Next.js 16 came out on 21 October 2025 and renamed middleware.ts to proxy.ts. A model trained before that date cannot know it, so it writes the old file name.

This is common. A study at ICSE 2025 found that models used deprecated APIs in 25% to 38% of completions in its tests, and bigger models used them more often. Another, at USENIX Security 2025, found that 19.7% of the packages models suggested did not exist at all. Newer models invent fewer: about 5% to 6% in a 2026 re-test. Fewer is still not zero.

The fix is a docs server: an MCP tool that fetches the current docs for the exact version you use. (MCP, the Model Context Protocol, is the standard plug that connects tools to an AI.) Context7 is the most starred (about 63,000 GitHub stars in October 2026); Microsoft Learn and GitMCP do the same for their docs.

# set up once
npx ctx7 setup

# then in a prompt
Add the auth check from the old middleware to proxy.ts for Next.js 16. use context7

# any GitHub repo's docs as an MCP server
github.com/owner/repo  ->  gitmcp.io/owner/repo

A docs server also puts outside text into your agent, and that is a risk. In August 2026 a flaw in Context7 up to version 2.1.2 let poisoned instructions tell agents to send out keys from .env files. Keep it updated, and never run an agent with production keys in its folder.

Put your team’s rules in one short file the agent reads first

Some knowledge is not in the code. “We never edit migrations by hand.” “Dates are Sydney time.” “That folder is dead, ignore it.” Agents read a rules file before they start: AGENTS.md (used by over 60,000 open-source projects and read by Codex, Cursor, Copilot’s agent, Gemini CLI and others), CLAUDE.md for Claude Code, or .cursor/rules.

This is the cheapest retrieval you can set up. A short one looks like this:

# AGENTS.md
## Project
Booking app for a renovation company. Next.js 16, Postgres.
## Commands
- Test: npm test
- Lint: npm run lint
## Rules
- Dates in Australian format, Sydney time.
- Never edit files in /migrations by hand. Ask first.
- Every new API route gets a test that tries to read another user's data.
## Where things are
- Bookings: src/bookings/
- Decisions and the reasons: docs/decisions/

Keep it short. Anthropic’s advice is “under 200 lines” per file, because “longer files consume more context and reduce adherence”. Put long how-tos in their own files and link them. The agent opens them only when it needs them. That is the idea behind Agent Skills.

Add the docs/decisions/ line even if the folder is small. A short note per decision, with the reason, stops the agent from “fixing” something your team chose on purpose.

Connect tickets and the wiki, read-only first

The why behind a change sits in Jira, Linear, Confluence, Notion or GitHub issues. All of them now have official MCP servers. With one connected, the agent can read the ticket and the acceptance criteria before it plans.

Start with read access only. In May 2025 researchers showed that a planted GitHub issue could make an agent copy private repo data into a public pull request. Whatever the agent reads can carry instructions. Read-only access and a person who approves every write keep that small.

Ask your codebase before you change it

Half of developers in Stack Overflow’s 2026 survey (49.7%) already use AI to understand or onboard to a codebase. Three questions that work well in a repo you do not know:

Where is a booking created, from the button to the database row? List the files in order.

If I change the price field from a number to a decimal type, what breaks? Search first, then list each place.

Which tests cover cancellation? Which paths have no test?

Ask for file names and line numbers in every answer. You can open them and check in seconds. In the same survey, 93% of developers named “source attribution” as something that makes them trust an AI answer. A file path is the simplest source.

When RAG makes the answer worse

Does the app I am building need RAG?

Sometimes the app itself needs to look things up: a support bot over your help pages, a search over a client’s contracts. Two facts help you decide. If all the text is under about 500 pages, Anthropic’s advice is to put it all in the prompt and skip retrieval. Above that, mix keyword search with vector search and add a reranker. In Anthropic’s tests, that mix with their contextual method cut failed lookups by 67%. In Menlo Ventures’ 2024 enterprise survey, 51% of companies used RAG, up from 31% a year before.

When a client app needs this, it is one of the things my team builds, with tests for the answers it gives. Checking those answers works like checking any code AI writes.

The numbers, by year

• 2025. Stack Overflow: about 31% used AI agents; 46% did not trust AI accuracy (Stack Overflow 2025). Among agent users at work, 34.4% used MCP servers.
• 2026. Stack Overflow: 66% use coding assistants or agents, 73% of them daily (Stack Overflow 2026). JetBrains, August 2026: 90% use AI coding agents at work every week. MCP SDKs: close to 500 million downloads a month, up from 97 million in December 2025.
• 2027 (my guess). Gartner expects 90% of enterprise engineers to use AI code assistants by 2028. My guess for 2027: developers stop choosing how the tool searches and spend that effort on what the agent reads: rules files, docs servers and decision notes.

Start this afternoon