Vector search, visualized.

How Salesforce Data 360 (formerly Data Cloud) turns documents into answers an agent can use: chunking, embeddings, vector search, and hybrid search.

Harmonization and identity resolution handle structured data: rows, fields, and keys. Most of what a company knows lives somewhere else, in policies, manuals, and help articles. Before an agent can answer from them, Data 360 has to cut them into pieces, turn each piece into numbers that capture its meaning, and find the right pieces when a question comes in.

Follow four short help documents from a fictional kitchen store through each step. Every choice you make here changes what the agent sees at the end.

1. Chunking

An agent can't read a whole manual for every question, and one vector can't hold a whole manual's meaning. So each document gets cut into chunks. Where the cuts fall decides what can be found later.

Chunking strategy

Max tokens per chunk:

In Data 360

When you create a search index, you pick what to index (text fields on a data model object, or files through an unstructured data model object) and a chunking strategy. Semantic-based passage extraction uses the structure in HTML, like headings, lists, and bold text, as boundaries. Window-based passage extraction uses blocks like paragraphs and line breaks, and falls back to sentences when there's no HTML. Either way, Data 360 splits the content into sentences and merges whole sentences into chunks up to the max tokens limit: 512 by default, estimated by word count. It doesn't cut a sentence in half. The chunks land in a chunk data model object, linked back to the source record.

The sample documents are tiny, so the max tokens range here is scaled down from the default of 512, and exactly where each strategy draws its boundaries is simplified.

↑ Back to top

2. Embeddings

Each chunk goes through an embedding model, which turns its meaning into a long list of numbers: a point in space. Chunks that mean similar things land near each other, even when they share no words. Returns sit near refunds. Error codes sit near cleaning, because both are about the blender itself.

Select a dot to read its chunk.

Real embeddings have hundreds or thousands of dimensions. This map flattens them to two, and the positions are illustrative, not the output of a real model. Change the chunking above and watch the dots move: a chunk's position comes from everything in it.

↑ Back to top

3. Vector search

A question gets embedded the same way, and becomes a point on the same map. Vector search returns the nearest chunks. That's why it understands "send it back" means "return," with no words in common.

Pick a question:

Nearest chunks

↑ Back to top

4. Hybrid search

Vector search has a blind spot: exact identifiers. "E17" and "E22" mean nearly the same thing to an embedding model (both are error codes), so it can't tell them apart. Keyword search is the opposite: it matches exact terms, and misses anything phrased differently. Hybrid search runs both and merges the rankings.

Pick a question:

Keyword weight:

Keyword Vector

Keyword half exact terms

Vector nearest in meaning

Hybrid both, fused

✓ has the answer marks chunks that contain what the question needs. Data 360 has no keyword-only index: the keyword column is the keyword half of a hybrid index, shown on its own so you can see what it adds.

In Data 360

A hybrid search index keeps a keyword index and a vector index side by side, runs a question against both, and fuses the results with a ranker. There are two: a linear fusion ranker (the default, tuned by Salesforce) and reciprocal rank fusion. This page uses reciprocal rank fusion, where the slider is Data 360's own keyword_weight setting: 0 to 1, default 0.5, with the vector weight set to whatever is left. You pass it per query, in the relevance settings of the hybrid_search() SQL function; it isn't a setting on the index itself. Reach for hybrid when your content is full of product codes, part numbers, error codes, or jargon. It consumes more credits than vector search alone, so check your contract for current rates.

↑ Back to top

What the agent sees

The retriever hands its top chunks to the model as grounding, and the model answers from those alone. If the right sentence didn't make it into a chunk that got retrieved, the model never sees it, however good the model is.

Try it: pick "Can I return a blender I've used?" and drag max tokens down. The rule and the sentence that limits it land in different chunks, and only one may come back.

↑ Back to top

The short version

Retrieval quality is decided long before the model runs. Chunk on the document's real structure so conditions stay attached to the rules they qualify. Use vector search when people ask in their own words, and hybrid when exact codes matter. Then test with real questions and read what actually got retrieved, not just the final answer.

The same idea runs through the other tools: harmonization and identity resolution decide what structured data an agent can trust. This is the unstructured half.

All content is fictional, the embedding map is illustrative, and the logic is simplified for learning. Keyword scoring (BM25) and the hybrid merge (reciprocal rank fusion) are computed live; Data 360's own scoring differs in detail. An independent tool, not affiliated with Salesforce. More Data 360 tools