Grounded Answers
Why a model answers from its training and what changes when you let it search: the same question asked twice, how web search grounding works, how to steer it, and the price and the rule.
Grounded Answers
Why a model answers from its training and what changes when you let it search: the same question asked twice, how web search grounding works, how to steer it, and the price and the rule.
You come away knowing what grounding and a training cutoff are, how the web search tool is switched on and steered with search context size, what it costs, and why a grounded answer still has to be verified.
Honest about how it was made: this film was drafted in Muse and then cleaned up and rebuilt with Claude. The film follows Meta's pricing page: web search grounding costs $2.50 per 1,000 search queries in addition to the request's token cost, not instead of it. Details of the course experiment the record does not support were cut, so the film says only that the grounded answer added retailers, shipping and price reality.
A note that opens the film Muse drafted this, Claude cleaned it up. This film began as a rough draft made with Muse. Muse runs on a simple model called Spark, and Spark makes a lot of mistakes, so Claude cleaned the draft up. The method is the point: draft in Muse, then finish in Claude or Codex. One kind of mistake we looked for: a draft Muse diagram of two big boxes and a check mark under a question about whether a non-technical person could use it, with both boxes empty. A diagram that shows nothing teaches nothing, so Claude replaced diagrams like this one.
Narrated by Liam, in for Bear. Narration is an AI voice; the film was made with AI tools.
Chapters 0:00 Muse note 0:28 The idea and the key terms 0:58 The Test 1:33 How It Works 2:40 Steering And Limits 3:30 The Price And The Rule 4:09 Recap 4:31 Your Turn
Your turn Paste this into Claude: I'm going to ask an AI model one question twice, once with web search off and once with it on. My question: [something whose answer changes with time, like the price of a part I need]. Before I run it, predict what the search-off answer will leave out and what the search-on answer should add. Then list what I should still verify myself. Then check two things yourself. Does it name things only a search could know? And does it tell you to verify the answer, not just trust it?
Sources shown on screen Meta's public docs at dev.meta.ai (search grounding, tool calling, pricing and rate limits), captured 2026-10-06. Narration follows a Muse course taught by Andrew Brown, named on screen. Nothing here is a product endorsement.
Part of the playlist Muse for Educational AI.
@NikBearBrown
What this video covers
- 0:00 — Muse note
- 0:28 — The idea and the key terms
- 0:58 — The Test
- 1:33 — How It Works
- 2:40 — Steering And Limits
- 3:30 — The Price And The Rule
- 4:09 — Recap
- 4:31 — Your Turn
Transcript
A lightly formatted transcript of the video, with timestamps so you can jump to any section on YouTube.
Muse note
[0:00] A note before we start. This film began as a rough draft made with Muse. Muse runs on a simple model called Spark and Spark makes a lot of mistakes. So Claude cleaned this draft up. The method is the point draft and Muse then finishing Claude or Codeex. Here is the kind of mistake we look for. A draft diagram from Muse. Two big boxes and a check mark under a question about whether a non-technical person could use it. Both boxes are empty. A diagram that shows nothing teaches nothing. So Claude replaced diagrams like this one.
The idea and the key terms
[0:28] Hello, this is Liam in forbear. A language model answers from what it learned in training and training stops on some date. Search grounding lets it look things up first. This lecture is about what that changes, what it can't reach, and what it costs. Six terms. Grounding. Tying an answer to live sources. Training cutoff. The date the model stopped learning. Tool. Something the model can call to get work done. Search query. One search the model runs. Citation. A link tying a claim to its source. and token, a small chunk of text, the unit you pay by. Andrew Brown,
The Test
[0:59] Who teaches a course on Muse, asked one question twice. Control arms for a Dodge Grand Caravan bought in Canada. Same question both times. The one difference was grounding, a toggle in the playground off, then on. With grounding off, the model answers from memory, and memory ends at the training cutoff. It can know what a control arm is. It can't check today's prices, which stores carry the part, or what shipping costs. With grounding on, the model searches first and answers second. Andrew Brown reports the grounded answer added three things: retailers, shipping, and price reality.
How It Works
[1:33] Meta's docs call grounding a built-in tool. Meta says there is no retrieval pipeline or search index for you to build. You add one tool to a request, and the model decides whether to search in code. Grounding is one extra argument. Take an ordinary request, add the web search tool, and the same call becomes a grounded one. It works on the responses of PI only, not chat completions. Turning the tool on does not force a search. The model checks the question for the capital of France. Meta says it typically answers from training. Ask who won the most recent Formula 1 race and it can search. When it does search, the response shows it. First comes a web search call a record that a search ran. Then comes the message with the answer text and a list of annotations. Each annotation is a citation. A citation holds the source's address and title, plus where in the answer the cited passage starts and ends. So each claim points to the page that supports it. Ask for the raw results, too, with include web search call results. That list holds every source the model considered. The citations are only the subset it cited. Compare the two and you see what the model saw. You can steer how much the
Steering And Limits
[2:41] Model reads. Search context size takes low, medium, or high. Low is the fastest and uses the fewest tokens. Medium is the balanced default. High is for complex questions that need more sources. You can also say where you are. Pass an approximate location and search leans toward that place. The country is a two-letter code. For question about buying in Canada, like Andrew Brown's, this is the field to reach for. Now, the limit Andrew Brown founding can't get through blocked product pages. If a retailer's page refuses the search, the model can't read it. So what that page knows never reaches the answer. Meta's own docs list the limits. Quality is still improving. Coverage is incomplete since not every source can be retrieved. It works best on focused questions. So treat the answer as a starting point and verify what you depend on. Now the price
The Price And The Rule
[3:31] Web search grounding costs $2.50 per thousand search queries. That is on top of the request's normal token cost. Do the arithmetic. A,000 searches cost $2.50. So, one search is a quarter of a cent. Then the tokens, $1.25 per million in,425 per million out. A grounded answer pays both. You can put a ceiling on it. Set max tool calls on a request and the models built-in tool calls such as web search, stop at that number. The minimum is one. So, here is a working rule of thumb. If the answer changes with time or depends on where you are, turn grounding on. If the question is timeless, leave it off and keep the fee.
Recap
[4:09] Let's recap with Claude. asked twice. The same question came back with more when grounding was on. Retailers, shipping, and price reality. Grounding is one built-in tool. The model decides whether to search and the answer carries citations. You can steer context size and location, but blocked pages stay out of reach and coverage is incomplete. And each search costs a quarter of a cent on top of tokens, so ground when the answer changes with time. Your turn. Paste this
Your Turn
[4:32] Into Claude. I'm going to ask an AI model one question twice. Once with web search off and once with it on. My question something whose answer changes with time like the price of part I need before I run it. Predict what the search off answer will leave out and what the search on answer should add. Then list what I should still verify myself. Then check two things yourself. Does it name things only a search could know? And does it tell you to verify the answer, not just trust it? Grounded answers at Nick Bear Brown.
