Formatting Retrieved Context for the Prompt

You could join retrieved chunks together with no labeling at all, and generation would still technically work. But without source boundaries, the model has no way to tell you which specific document backed a specific claim — and neither do you, when debugging a wrong answer. Numbering sources is a small formatting step with an outsized payoff for trust and debuggability.

Real RAG systems often go further than plain numbering — adding metadata like document titles, timestamps, or URLs next to each source. The core pattern stays the same as what you just built: retrieve, then format each result as a clearly delimited, labeled unit before it ever reaches the prompt.

From Chunk IDs to Prompt Text. similarity_search() returns ids like 'pto-policy' — useful for your code, meaningless to the LLM. Before generation, you need to turn retrieved ids into real, labeled text the model can read and even cite back to the user.

Build the Context Formatter. Finish format_context(): for each retrieved id, look up its full text and append a labeled line like '[Source N] <text>' to lines. Numbering the sources isn't cosmetic — it's what lets the model cite exactly which source it used in its answer.

Why label each retrieved chunk with a source number like '[Source 1]' instead of just concatenating the raw text together?

• →It lets the generation step (and the model's answer) reference exactly which retrieved source backs a specific claim, which is essential for user trust and debugging wrong answers.
• →It reduces the total number of tokens sent to the model.

A Real, LLM-Ready Context String. You can now go from a raw query all the way to a clean, labeled context string. Next lesson closes Module 2: combining similarity_search() and format_context() into a single retrieve() function — the one piece your chatbot will actually call.