WEBVTT

00:00:00.000 --> 00:00:02.420
Your AI says you have thirty days to

00:00:02.420 --> 00:00:04.880
return a camera. It even links to the

00:00:04.880 --> 00:00:09.160
shop's policy. Great. Except that policy changed

00:00:09.160 --> 00:00:13.620
last month. Now it's fourteen days. This is

00:00:13.620 --> 00:00:17.060
a made-up shop, but a useful question: how

00:00:17.060 --> 00:00:19.360
can an answer have a source and still

00:00:19.360 --> 00:00:23.660
be wrong? Meet RAG: retrieval-augmented generation.

00:00:24.820 --> 00:00:29.120
Terrible name. Pretty sensible idea. Instead of relying

00:00:29.120 --> 00:00:31.600
only on what a model learned during training,

00:00:32.240 --> 00:00:35.440
an application searches outside information and gives

00:00:35.440 --> 00:00:38.060
relevant passages to the model before it answers.

00:00:39.320 --> 00:00:42.360
Think of an open-book exam. The model still

00:00:42.360 --> 00:00:45.140
has to write the answer. Now it has

00:00:45.140 --> 00:00:47.360
some pages to work from. Here's our camera

00:00:47.360 --> 00:00:50.360
question going through a simple version. First, search

00:00:50.360 --> 00:00:53.180
a collection of documents. That might use keywords,

00:00:53.180 --> 00:00:57.400
similarity in meaning, or both. Next, bring selected

00:00:57.400 --> 00:01:00.420
passages into the model's working context, alongside

00:01:00.420 --> 00:01:04.080
the question. That's the augmentation part: adding

00:01:04.080 --> 00:01:06.480
information for this answer, not automatically

00:01:06.480 --> 00:01:10.400
retraining the model. Finally, the model generates a

00:01:10.400 --> 00:01:13.760
reply. The application can attach references so you

00:01:13.760 --> 00:01:17.320
can inspect its sources. Search. Supply.

00:01:18.220 --> 00:01:20.860
Answer. But look at the page our imaginary

00:01:20.860 --> 00:01:25.120
system picked. It's the old returns policy. Thirty

00:01:25.120 --> 00:01:28.520
days. The current one says fourteen, but it

00:01:28.520 --> 00:01:30.420
never made it into the information given to

00:01:30.420 --> 00:01:33.820
the model. The answer follows the retrieved page.

00:01:34.540 --> 00:01:37.580
The citation points to that page. And you

00:01:37.580 --> 00:01:40.740
still miss the return deadline. A source link

00:01:40.740 --> 00:01:43.600
tells you where to look. It doesn't certify

00:01:43.600 --> 00:01:46.540
that the source is current, applicable, or correctly

00:01:46.540 --> 00:01:50.000
interpreted. Even with the right page, a model

00:01:50.000 --> 00:01:52.740
can misread it or add something it doesn't

00:01:52.740 --> 00:01:55.960
support. So the repair isn't just, "Get a

00:01:55.960 --> 00:01:59.520
smarter model." For this example, mark the old

00:01:59.520 --> 00:02:02.660
policy as superseded and keep it out of

00:02:02.660 --> 00:02:05.960
current answers. Make the current, approved policy

00:02:05.960 --> 00:02:08.840
available to search, and check that the retriever

00:02:08.840 --> 00:02:12.500
actually selects it. Then inspect the answer against

00:02:12.500 --> 00:02:15.920
the passage. If the documents conflict or don't

00:02:15.920 --> 00:02:18.520
contain the answer, design the system to say

00:02:18.520 --> 00:02:21.740
so, and test whether it does. That's why

00:02:21.740 --> 00:02:24.820
RAG can be useful for changing manuals, support

00:02:24.820 --> 00:02:28.380
information, or private documents. But access controls

00:02:28.380 --> 00:02:31.440
still matter. A document existing somewhere doesn't

00:02:31.440 --> 00:02:33.280
mean every user should be able to retrieve

00:02:33.280 --> 00:02:36.080
it. Next time an AI answers "from your

00:02:36.080 --> 00:02:39.760
documents," check three things. Which passage did it

00:02:39.760 --> 00:02:43.300
use? Is that the right version? Does it

00:02:43.300 --> 00:02:47.080
support this answer? RAG brings information to the

00:02:47.080 --> 00:02:49.620
model. Checking the evidence is still part of

00:02:49.620 --> 00:02:52.840
the job. Keep exploring with Archive of Worlds.

00:02:53.460 --> 00:02:56.160
Subscribe for more clear explanations, and visit our

00:02:56.160 --> 00:02:57.180
website for the sources.