Skip to content

Search and Retrieval — Junior

At junior level, focus on this question:

Given a question that should be answerable from a repo or a set of documents, can you use grep/regex/glob to find the exact relevant lines instead of pulling in whole files?

grep is a tool, not a fallback

For our data-analyst agent working against a dbt repo, "which model defines gmv_daily?" is answered by one command, not by loading every .sql file into context:

rg -n "def gmv_daily|model.*gmv_daily" --type sql
  • rg/grep do exact and regex matching — they find precisely what you asked for, with no approximation and no missed synonyms if the term is right.
  • Returning line numbers and a small context window (-A 3 -B 3 = 3 lines after/before) costs a fraction of the tokens that returning the whole file would cost, and gives the model exactly the span it needs.
  • glob (e.g., **/*.sql) narrows which files to look in before searching their contents — combine narrowing with searching rather than searching everything.
flowchart LR A["Question with an exact term<br/>'gmv_daily'"] --> B["glob: narrow file set"] B --> C["grep: find exact matches"] C --> D["Return line + small context window"]

When exact match is the right tool

  • Looking for an identifier: a column name, function name, error code, config key, order ID.
  • Looking for where something is defined, not "what does this concept mean."
  • The corpus is code, logs, or config — text where meaning is tied to exact tokens, not paraphrase.

When exact match falls short

  • The question uses different words than the document ("revenue drop" vs. a doc that says "GMV decline") — a plain grep for "revenue drop" finds nothing even though the relevant doc exists. This gap is what BM25 and later RAG address.

Trace one query by hand

For "find every place refund_amount is validated" in a sample repo:

  1. Write the exact grep/rg command, including any file-type or path filter.
  2. Predict how many matches you expect before running it.
  3. Run it. If it returns hundreds of matches, that's a signal your pattern is too broad — narrow it (a file glob, a more specific regex) rather than returning all of it to the model.
  4. Write the final line-number + snippet set you'd hand to the model — not the whole files.

Comprehension check

  • Why is returning a line number and a small snippet better than returning the whole file, for both accuracy and token cost?
  • Give one example of a question where grep is the obviously correct tool, and one where it obviously is not.
  • What's wrong with a search that returns 5,000 matches with no filtering, and what's one way to fix it?
  • What does glob narrow, and why narrow before searching contents?