Retrieval Context and the AI Evidence Bridge

StretchSearch isn't only a search UI — it's the retrieval layer that grounds AI. When another StretchSuite app (like StretchGPT) needs to answer a question with first-party facts instead of guesses, it asks StretchSearch for retrieval context and an evidence package. This article explains both.

Retrieval context

Retrieval context is the set of evidence chunks most relevant to a query. Given a question, StretchSearch returns the page chunks that best match, scoped to whatever you specify — a project, a crawl job, or a single page. These chunks are the grounding material an AI model reads before it answers.

When semantic re-ranking is enabled, StretchSearch refines the chunk order using embeddings on a GPU node for better relevance. The feature is fail-safe: if it's turned off or the embedding node is unreachable, StretchSearch simply returns the keyword-ordered chunks unchanged, so retrieval never breaks.

The evidence package (the AI bridge)

The evidence package is the bundle StretchSearch hands to AI surfaces. For a given query it combines:

• Retrieval context — the relevant evidence chunks (up to a sensible cap).
• Ranking overview — the current ranking-preview summary and status.
• Ranking previews — a slice of top ranked candidates.

Together these give a model both the raw evidence to quote and the ranked signals to weigh, so answers are grounded and citable.

Request retrieval context

1. From the retrieval surface (or a calling app), provide a query.
2. Optionally scope it with project, crawl job, or page.
3. Set a limit for how many evidence chunks to return (up to 25).
4. StretchSearch returns the ranked evidence chunks, flagged if they were semantically re-ranked.

Request an evidence package

1. Provide the query and any scope (project, crawl job, page).
2. Set the limit for evidence chunks.
3. StretchSearch returns the combined package — retrieval context plus the ranking overview and top previews — ready for an AI grounding step.

How the suite uses it

• StretchGPT and other AI answer surfaces call the bridge so their responses cite your indexed content.
• StretchBAS can pull grounded evidence to support audit narratives.
• Any app that needs "answer with proof" retrieves the package rather than reasoning from thin air.

Because the package is built from the same deterministic candidates and evidence references used elsewhere, the citations an AI produces trace back to real, versioned sources.

A realistic example

A client asks Northwind's agency, via a StretchGPT-powered assistant, "Does the site clearly explain your roasting process?" The assistant queries the evidence bridge scoped to the Northwind Coffee — Q3 SEO project. StretchSearch returns the most relevant page chunks about roasting, semantically re-ranked, plus the top-ranked entity/phrase candidates. The assistant answers with a summary and links back to the exact pages — grounded, not guessed.

Tips

• Scope retrieval to a project or crawl job for sharper, on-topic evidence — unscoped queries pull from everything indexed.
• Keep the limit modest (a handful of chunks) for focused answers; raise it only when you need broad coverage.
• If semantic re-rank isn't available, retrieval still works — you just get keyword-ordered chunks. Answers remain grounded.
• Pair the bridge with evidence references when you need to show exactly which versioned source backed an answer.

FAQ

Does the AI bridge send my data to an external model? StretchSearch supplies grounding context to StretchSuite's own AI surfaces. The retrieval layer's job is to return your indexed evidence; how a given app uses it is governed by that app.

What happens if the semantic re-ranker is down? Retrieval fails safe and returns the keyword-ordered chunks. You lose the embedding refinement, not the results.

Can I ground answers on a single page? Yes — pass a page scope and StretchSearch limits retrieval to that page's chunks.