---
title: "RAG is for finding. Full context is for deciding."
description: "Retrieval finds the closest match; judgement needs the whole picture. Why I send the full CV instead of top-k chunks, and when I still use RAG."
url: https://seo7.es/en/blog/inteligencia-artificial-blog/rag-for-finding-full-context-for-deciding
canonical: https://seo7.es/en/blog/inteligencia-artificial-blog/rag-for-finding-full-context-for-deciding
lang: en
type: BlogPosting
---

# RAG is for finding. Full context is for deciding.

> Retrieval finds the closest match; judgement needs the whole picture. Why I send the full CV instead of top-k chunks, and when I still use RAG.

Page: https://seo7.es/en/blog/inteligencia-artificial-blog/rag-for-finding-full-context-for-deciding

## Summary for AI

Retrieval finds the closest match; judgement needs the whole picture. Why I send the full CV instead of top-k chunks, and when I still use RAG. The advice in 2026 is settled: chunk your documents, embed them, retrieve top-k, feed those to the model. I built a system that deliberately does none of that. Every query gets the expert's entire CV: 69 projects, publications, education, languages.

## Questions and answers

### Why does the tender matching system give the model the whole résumé?

Each request gets the expert's entire résumé: 69 projects, publications, education and languages, about 34,000 characters. The system's job is judging whether the person fits, and to judge, the model needs the whole picture, including what the résumé lacks.

### How does retrieval differ from exclusion?

Retrieval searches the corpus for what looks most like the query. Exclusion searches for what disqualifies. These are opposite operations: the first needs the strongest matches, the second the weakest link, which by definition looks least like the rest.

### How many UN tender sources does the system crawl every day?

Twenty-eight. The system normalizes and deduplicates thousands of procurement notices and, for each tender, answers one question: is this expert a fit.

### Why does cosine similarity fail at selecting tenders?

A Public Financial Management tender and a dozen projects by a public finance expert give very high similarity. But the tender may require an ecologist, mentioned once on page five, or a national consultant with the local language. RAG will return the finance projects and decide the match is strong.

### Why doesn't the disqualifying detail surface in top-k?

It is never the semantically closest fragment: that is exactly why it disqualifies, it's the part that breaks the pattern. A retrieval system is optimized to surface similar things and hides just this kind of detail.

### What does a résumé lose when split into chunks?

The facts stay, the shape is lost. Three chunks about public finance show the model a finance expert. All 69 projects show where he actually worked, which languages he speaks, what seniority level he works at and what he has never done.

### Why doesn't RAG notice missing experience in a document?

There is no chunk saying "this person has no experience in ecology". You can only learn it by reading the whole document and noticing that such experience isn't there. RAG tells you what the document contains, but not what it lacks.

### How much does it cost to give the model 9,000 tokens for every tender?

About 34,000 characters is roughly 9,000 tokens per judgment, multiplied by hundreds of tenders a day. In practice AI and proxies together cost a few dollars a month, and the dashboard calculates the cost of a run and the price of each tender found.

### How does the prefilter cut model costs?

The cheap work is done by code with no model calls: stop words, topic, territory, language, deadline, expiry. The model only sees candidates that survived these rules, so there are few expensive judgments with the full résumé.

### Why does the full résumé cache well in the prompt?

The résumé is byte-for-byte identical on every call. The same prefix repeats again and again, which is the ideal shape for prompt caching: the repeated part of the prompt costs less.

### In what form does the model return its verdict on a tender?

As structured JSON that is logged in full: fits, short_reason, long_reason, country and tor_attached. For example, fits: false with the reason "a national consultant is required, local language is mandatory". The model judges the substance of the role, not its title.

### When is RAG still needed?

When the corpus is unbounded: a knowledge base, documentation or support history that grows endlessly. When latency matters. And when the task really is search: find a fragment and answer from it. The author uses RAG on other projects for exactly these reasons.

### What are the limits of long context?

Models retain the middle of a long prompt worse than the beginning and the end. 34 thousand characters fit within what current models handle well. At 300 thousand, the author would most likely build a hybrid of search and full context.

### How do you decide between RAG and full context?

Ask what the model must do with the document. Finding something means retrieval: the corpus can be huge. Judging something means giving it everything, because irrelevance is half the verdict. Candidate matching, contract risk and tender participation are judgment tasks.

## Other languages

- RU: https://seo7.es/ru/blog/inteligencia-artificial-blog/rag-for-finding-full-context-for-deciding
- ES: https://seo7.es/es/blog/inteligencia-artificial-blog/rag-for-finding-full-context-for-deciding

## schema.org markup

```json
{
  "@context": "https://schema.org",
  "@type": "BlogPosting",
  "name": "RAG is for finding. Full context is for deciding.",
  "description": "Retrieval finds the closest match; judgement needs the whole picture. Why I send the full CV instead of top-k chunks, and when I still use RAG.",
  "url": "https://seo7.es/en/blog/inteligencia-artificial-blog/rag-for-finding-full-context-for-deciding",
  "mainEntityOfPage": "https://seo7.es/en/blog/inteligencia-artificial-blog/rag-for-finding-full-context-for-deciding",
  "inLanguage": "en-GB",
  "headline": "RAG is for finding. Full context is for deciding.",
  "isPartOf": {
    "@type": "WebSite",
    "name": "SEO7.ES",
    "url": "https://seo7.es/"
  },
  "publisher": {
    "@type": "Organization",
    "name": "SEO7.ES",
    "url": "https://seo7.es/"
  }
}
```