Grounding and Verification

After Hallucination, Change the System

The previous page established a crucial fact: a mechanically correct generation process can produce a false statement. Telling the same generator to “be more accurate” does not by itself create evidence. A stronger system can bring relevant evidence into the inference and then test important claims against evidence or computation.

Three Different Jobs

Where Does Grounding Come From?

The Model Does Not Magically See the Source

A document sitting on an SSD, a web page on another server, or a row in a database is not automatically part of the LLM's current context.

Something in the surrounding system must retrieve or otherwise provide the information.

Retrieval-Augmented Generation — RAG

A common pattern is called Retrieval-Augmented Generation.

The retrieval system and the language model perform different jobs. Retrieval finds candidate evidence. The LLM uses that evidence while generating.

A Tiny Example

Suppose we ask about a machine whose current configuration is stored in a technical inventory.

The current fact came from the inventory record, not from hoping the model's training happened to contain it.

Retrieval Is Selection

For a large document collection, the system usually cannot stuff everything into the context. It has to choose useful pieces.

This creates a new failure point: the generator cannot faithfully use evidence that the retrieval system failed to find.

Embeddings Return — But With a Different Job

Retrieval systems often use vector embeddings to represent passages and queries so that semantically related material can be found even when the exact words differ.

Do not automatically equate these retrieval embeddings with the token embedding lookup inside the LLM. Both use vectors, but they can be produced by different models and serve different purposes.

Vector Search Is Not Truth Search

A similarity search asks roughly:

It does not inherently ask:

Retrieved material still needs source quality, relevance and interpretation.

Grounding Can Fail

Citation Is Provenance, Not Magic

A citation can help a reader trace a claim to its supporting source.

But merely attaching a citation does not prove the cited source actually supports the sentence. Citation quality itself can be evaluated.

Now Add Verification

Grounding supplies evidence before or during generation. Verification can examine the candidate result afterward—or sometimes during a multi-step process.

Different Claims Need Different Verifiers

There is no single universal “truth API.” Verification depends on what kind of claim is being made.

Example: Let the Calculator Own the Arithmetic

The language model can explain the procedure while deterministic arithmetic independently checks the numerical result.

Example: Let the Document Own the Fact

The source, not the model's confidence, is the authority for that claim.

Authority Matters

Two sources can disagree. Grounding therefore has another dimension:

A random copied page and the system of record should not automatically receive equal weight.

Provenance: Where Did This Fact Come From?

For serious verification we want more than the text of a claim. We may want metadata:

This should look familiar. It is the same evidence discipline that appeared in WITNESS.

The Chain of Evidence

If we preserve this chain, an investigator can ask where an error entered rather than merely observing that the final answer was wrong.

Grounding Does Not Rewrite the Weights

Ordinary RAG changes the information available in the current context. It does not normally retrain the model or permanently write the retrieved document into the weights.

Verification Does Not Require the LLM to Judge Itself

A model can critique or review its own answer, and that can sometimes help. But a stronger architecture can use independent mechanisms:

The verification mechanism should fit the claim and, when possible, provide evidence independent of the original generation.

What Happens When Verification Fails?

A well-designed system has choices besides inventing another answer.

Confidence Is Not Verification

A model's apparent certainty is a property of its generated language, not an independent measurement that the proposition is true.

Grounding Is a System Property

Return to the stack:

This is why LLM and AI assistant system must not be treated as synonyms.

WITNESS Has a New Question

If an AI system gives a grounded answer, WITNESS would not merely want the final sentence. It could ask:

Now we can reconstruct not only what the model said, but part of the evidence path that allowed the system to say it.

HAAW Checkpoint

Grounding

Providing relevant external evidence or authoritative information for the current inference.

Retrieval

Finding candidate information from documents, databases, search systems or other stores.

RAG

Retrieval-Augmented Generation: retrieve relevant information, place it in context, then generate using it.

Retrieval embedding

A vector representation used to help search or rank semantically related material; not automatically the same as the LLM's internal token embeddings.

Vector search

Retrieval based on numerical similarity in an embedding space. Similarity does not itself establish truth.

Verification

Testing a generated claim against evidence, computation, tools, constraints or authoritative records.

Provenance

Information describing where evidence came from, its version or time, and how it was obtained or transformed.

Citation

A reference connecting a claim to a source so its support can be inspected; a citation itself is not proof that support is valid.

Authority

The degree to which a source is appropriate and trustworthy for the particular claim being checked.

Abstention

Declining to make a factual claim when available evidence is insufficient.

Central lesson

Do not ask plausibility to do the job of evidence. Retrieve evidence, preserve its provenance, generate with it, and verify important claims with mechanisms suited to the claim.