Training stops; your data does not.
A model's facts end at training time and never include your private, fast-moving accounts. The deal that moved on Tuesday is simply not in there.
RAG over CRM, email, and multiple sources
A model answers from what it learned in training. Retrieval-augmented generation lets it look things up first, then answer from what it found. The result can be current, specific to your accounts, and traceable back to the source of every claim.
Look up the evidence before answering
QuestionWhat is still blocking the Acme pilot?
Retrieved passages
1 · Call note
Security sign-off is still pending.
2 · Email
Jon needs the SSO setup guide.
Answer from those sources
Security approval is the remaining blocker.1 Jon is waiting for the SSO guide before the review can move forward.2
A closed-book exam, turned open-book.
A language model alone answers from frozen, memorized knowledge. RAG lets it consult a live, updatable set of sources at the moment it answers.
Retrieval-augmented generation is two steps. Look up the material that bears on a question, then write the answer from it. The model still writes in natural language. The facts come from passages it just pulled, so the answer can be current, specific to you, and checkable, instead of a confident guess.
The paper that named the pattern, Lewis et al., 2020, frames it as two memories working together: the model's learned language ability, and a searchable index of documents it can consult. They call these parametric and non-parametric memory. The plain version: what the model knows by heart, and what it can look up.
Keep one running example in mind through the page. A question asked over a messy mix of email, CRM rows, call notes, and shared docs, the way a real account actually lives.
The prompt gains a source lookup
Without retrieval
- Question
Ask the model from its existing context.
- Answer
Rely on available prompt content and learned patterns.
With retrieval
- Question
Search the relevant source collection.
- Evidence
Include selected passages in the prompt.
- Answer
Respond using those passages.
The looking-up is the hard part.
Answering fluently is easy. Answering from the right evidence, in a way you can check, across messy and access-controlled sources, is a stack of separate problems.
Fluent and wrong is still wrong.
Without evidence to stand on, a model can state something untrue in clean prose. Grounding the answer in retrieved passages is the main defense.
People ask the same thing many ways.
"Revisit pricing" and "can we talk about the rate" share no words. Exact-match search alone misses the connection, because it is matching letters, not meaning.
Every source is a different shape.
A CRM field, an email thread, a slide, and a call note carry different structure. Pulling one clean answer across all of them is the real work.
Pointing to the proof is its own problem.
Producing an answer is easy. Producing one where each claim points back to the exact passage that supports it is a separate, measurable task.
Knowing everything is not seeing everything.
Retrieval has to respect who is allowed to see what, before an answer forms. And the index has to stay fast and current as documents change.
A question finds its evidence.
The pipeline, one idea at a time. Each beat names the real thing one layer down, and links the paper behind it.
1Embeddings
Text becomes a point in space.
Each chunk of text is turned into a vector, a list of numbers that encodes its meaning, so passages that mean similar things sit near each other. Search becomes "find the nearest points," which catches paraphrases that keyword search misses.
Paraphrases can land near each other
Pricing-related passagesUnrelated venue passage
- Query
- Revisit pricing.
- Nearby meaning
- Can we talk about the rate? · Discuss the quote.
2Dense retrieval
Learn to put a question near its answer.
A dual-encoder learns to place a question and its answer-bearing passage close together in that space. Dense Passage Retrieval (Karpukhin et al., 2020) reports retrieval-accuracy gains of roughly 9 to 19 points over classic keyword search (BM25). This is the engine of "find by meaning."
For precision at scale, ColBERT (Khattab and Zaharia, 2020) keeps word-level detail in the comparison while pre-computing document representations, so meaning-search stays both precise and fast across large collections.
Retrieve candidates before generating an answer
- Encode the query
Create a query representation.
- Search
Find candidate passages in the index.
- Rerank
Optionally reorder candidates using a richer comparison.
- Generate
Supply selected evidence to the answering model.
3Hybrid retrieval
Two ways to find the right passage.
Meaning-search and exact-match search are complementary. Meaning-search catches paraphrases. Exact-match catches codes, SKUs, names, and IDs that have to land letter-for-letter. The strong systems run both and merge the candidates, so they miss less. Documents are split into passages and stored in a vector index first, so the relevant pieces can be fetched fast.
Combine semantic and exact-match candidates
Meaning search
Example input
“Can we talk about the rate?” matches a pricing question.
Exact-match search
Example input
“TS-999” matches the specific error code.
Merge and rerank
Combine candidate sets before choosing the final evidence.
4Multi-source fusion
Many passages, one answer.
Real questions need several passages from several places. Fusion-in-Decoder (Izacard and Grave, 2021) showed a generator reading many retrieved passages at once and fusing them into a single answer, rather than leaning on any one source. A question spanning email, CRM, notes, and docs resolves into one grounded answer, not ten open tabs.
Keep the source beside each retrieved fact
| Source | Relevant passage |
|---|---|
| CRM | Deal is in negotiation. |
| Revised quote requested by Friday. | |
| Meeting note | The CFO must approve. |
| Document | Current renewal terms. |
5Grounding and citations
Footnotes you can check.
Good systems do not just answer. They attach each claim to the passage that supports it, like footnotes in a research paper, so a reader can verify it. The ALCE benchmark (Gao et al., 2023) defines and measures exactly this: text generated with citations you can check. It turns "trust me" into "here is the receipt."
Cite each claim where it is made
- Revised quote requested
- Email from the client, latest thread.
- CFO approval required
- Meeting note, approval discussion.
- Renewal terms attached
- Current renewal document.
6Permission-aware retrieval
Scoped before it answers.
In a real company, knowing everything cannot mean everyone sees everything. The candidate set is filtered to what the asker is cleared to see before the answer is formed. A source outside someone's access simply never enters retrieval, so it cannot reach the answer. This is access-scoping of what retrieval can reach, described as a principle.
Apply access rules before building context
| Candidate source | Decision |
|---|---|
| Assigned account notes | Include if permitted |
| Shared team document | Include if permitted |
| Another restricted team’s deal | Exclude |
Authorized evidence only
The answer is generated from the allowed subset.
Current, specific, and checkable.
When retrieval does its job, the answer changes character. Less guesswork, more evidence.
The system can answer from this week's data and from your own material, with no retraining. Update the index, not the model, and the answers update with it. Every claim can carry the passage it stands on, so a reader can check it at a glance. A question spanning email, CRM, notes, and docs resolves into one grounded answer. And because retrieval is scoped, broad knowledge and tight access control coexist: each person gets answers only from what they are allowed to see.
A useful answer makes its limits visible
- Current
Identify the date and version of the evidence.
- Checkable
Connect claims to supporting passages.
- Scoped
Use only sources allowed for this request.
- Honest about gaps
State when the evidence does not answer a detail.
From flat passages to maps, plans, and proofs.
The field is moving past a single look-up toward richer structure, multi-step search, and stronger verification.
A map, not just a pile of passages.
GraphRAG (Edge et al., 2024) builds a graph of entities and relationships, plus cluster summaries, so the system can answer broad, whole-corpus questions that flat passage search struggles with.
Knowledge graphs →Search, read, then search again.
Systems that plan across steps, retrieving more than once, rather than looking up a single set of passages and stopping. Long context and retrieval are also being combined: when to read more, and when to go and fetch.
Check that every sentence is supported.
Stronger attribution and automatic checking that each claim is actually backed by its cited source. Retrieval is also reaching beyond prose, into tables, slides, and images.
What this looks like as an everyday outcome.
This family of techniques is how the industry gives an assistant current, sourced, permission-scoped answers across many places at once. Here is what that feels like for a seller.
Ask about an account in plain words and get an answer you can check, drawn from email, calendar, CRM, call notes, and your team's docs. One picture of the account instead of ten open tabs. Every answer can show where it came from, so you trust it at a glance. And you only ever get answers from what you are allowed to see.
AI REV Voice runs on Prism, our own model, built for sales conversations. This page explains the field, not the plumbing underneath any one product.
See the account picture and answers with sources as product behavior on Deep context, meet the model on Prism, or hear it answer out loud on the demo.
Answer the account question from its evidence
Example input
Where did we land with Acme, and what do I owe them?
Prepare the revised quote by Friday.
The latest email requests it. The meeting note says the CFO now approves the proposal.
| Detail | Source |
|---|---|
| Revised quote by Friday | Client email |
| CFO approval | Latest meeting note |
The papers behind each idea.
01
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
arxiv.org/abs/2005.11401 ↗
02
Dense Passage Retrieval for Open-Domain Question Answering
arxiv.org/abs/2004.04906 ↗
03
ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT
arxiv.org/abs/2004.12832 ↗
04
Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering
arxiv.org/abs/2007.01282 ↗
05
Enabling Large Language Models to Generate Text with Citations
arxiv.org/abs/2305.14627 ↗
06
From Local to Global: A Graph RAG Approach to Query-Focused Summarization
arxiv.org/abs/2404.16130 ↗
Keep reading.
Knowledge graphs
Where retrieval is heading: a map of people, deals, and documents that answers the broad questions no single passage holds.
Read the explainer →Secure enclaves
The boundary side of the story: how sensitive computation stays inside a protected space, in plain terms.
Read the explainer →Speech to text
Where a lot of this material starts: turning a spoken call into clean, searchable text in the first place.
Read the explainer →Talk. Done.
Reading is one thing. Hand a phone number over and ask an account a question out loud.
No app required, just your phone number.