Grounding Answers that can show where they came from.
Short answer. Grounding means tying an AI model’s answer to specific, checkable sources, such as passages from the office’s own documents, rather than letting it answer from memory. The usual method is retrieval: find the relevant passages, give them to the model, and require the answer to cite them. Grounding lets a person verify an answer in seconds, but it does not make the answer correct. In the ALCE benchmark, even the best models lacked complete citation support 50% of the time on one dataset, so a grounded system should show the quoted passage and not only a reference number.
01
How it works
A grounded system has five steps. It finds the passages that bear on the question, usually with embeddings and a vector search. It puts those passages in the model’s context. It instructs the model to answer only from them. It asks for a citation for each claim, ideally a short quote. And it lets a person or a program check the citation against the original text.
The idea predates today’s assistants. The 2020 paper that introduced retrieval-augmented generation combined a model’s built-in knowledge with a searchable index of text, and noted that giving provenance for a model’s decisions and updating its knowledge were open problems that this approach could address. Tool results, such as a database lookup or a calculator, also ground an answer.
02
Levels of grounding
Level
What the user sees
What it gives you
Ungrounded
An answer from the model’s memory
Nothing to check against
Retrieved context
An answer plus a list of documents consulted
Shows what was available, not what was used
Cited passages
An answer with the exact quoted passage for each claim
A person can check each claim in seconds
Verified quotes
Software confirms the quotes exist in the source text
Catches invented quotes
Confirmed
A person approves before the result becomes a record
The highest assurance, at the cost of time
03
How to measure it
Citation support: does each cited passage actually back the claim? The ALCE benchmark scores citation quality as well as fluency and correctness.
Retrieval hit rate: for a set of real questions, does the right passage appear among the passages given to the model?
Faithfulness: does the answer say only what the passages say?
Abstention: when the documents do not contain the answer, does the system say so? In ANVE’s published harness a 4B-class model declined 5 of 5 such questions.
04
Where grounding fails
The wrong passage is retrieved, so the answer is faithful to the wrong text.
The right passage is retrieved but misread, for example a figure taken from the wrong column.
The citation points to a passage that does not support the claim.
The documents are out of date or contradict each other, and the system does not say so.
Permissions are ignored, so a user is shown a passage from a file they may not read.
05
Where ANVE stands
ANVE’s design goal is that every answer is cited, every action is owned and every model call is logged. In its document question answering, the source passage is shown beside the answer. Each meeting action item must quote the transcript line it came from, and a person confirms the owner before it is approved. In the agent loop, a claim made without a successful tool call is challenged.
06
Questions people ask
Is grounding the same as RAG?
RAG, retrieval-augmented generation, is the most common way to ground an answer. Grounding is the goal of tying answers to sources, and tool calls or database lookups can also do it.
Does grounding stop hallucination?
It reduces it but does not stop it. The model can still misread a passage or cite one that does not support its claim, which is why the quoted passage should be shown.
What should a good citation look like?
The document name, the page or section, and the exact words relied on. A reference number alone does not let a person check quickly.
Does grounding need the internet?
No. It can use only the office’s own documents. Web lookup is optional, and in ANVE’s research maker it happens only when asked.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
