WEBVTT

1
00:00:08.570 --> 00:00:13.380
When it comes to AI,
retrieval-augmented generation, or RAG,

2
00:00:13.380 --> 00:00:17.668
is one of the most fundamental building
blocks of any intelligent application.

3
00:00:18.505 --> 00:00:21.060
Most people think of RAG as a bolt-on:

4
00:00:21.060 --> 00:00:23.994
something that fetches
data from a document store,

5
00:00:23.994 --> 00:00:27.808
shoves it into a large language model,
and hopes for the best.

6
00:00:28.164 --> 00:00:32.195
But inside ServiceNow,
RAG is not an afterthought.

7
00:00:32.552 --> 00:00:37.533
It's embedded right into the fabric
of the platform, through AI Search.

8
00:00:38.370 --> 00:00:44.253
The words around it can be daunting:
indexing, retrieval, chunking, scoring,

9
00:00:44.253 --> 00:00:46.368
re-ranking and embeddings.

10
00:00:46.725 --> 00:00:50.054
But once you understand
how ServiceNow structures them,

11
00:00:50.054 --> 00:00:53.283
you can start to tune RAG
for your own use cases.

12
00:00:54.503 --> 00:00:58.104
At the core of ServiceNow's
RAG is semantic search,

13
00:00:58.104 --> 00:01:01.292
built on a set of linked
configuration tables.

14
00:01:02.129 --> 00:01:04.547
At the top is the search application:

15
00:01:04.547 --> 00:01:09.882
the experience itself, such as Now
Assist, Virtual Agent or the portal.

16
00:01:10.239 --> 00:01:14.165
The search profile decides
which data sources it uses,

17
00:01:14.165 --> 00:01:16.905
and how it ranks and filters the results.

18
00:01:17.262 --> 00:01:22.706
A search source is a single table,
such as knowledge articles or incidents.

19
00:01:23.062 --> 00:01:27.324
An indexed source marks it
as enabled for semantic search.

20
00:01:27.681 --> 00:01:31.621
The semantic index config
controls the vector embeddings,

21
00:01:31.621 --> 00:01:33.709
the heart of semantic search.

22
00:01:34.066 --> 00:01:39.061
And semantic fields are the fields
whose text is embedded into vectors.

23
00:01:39.898 --> 00:01:44.784
Together, they convert ordinary
records into vectorised meaning.

24
00:01:45.141 --> 00:01:47.573
Indexing runs from the bottom up,

25
00:01:47.573 --> 00:01:50.911
from each semantic field
to the search application.

26
00:01:52.131 --> 00:01:55.351
When a user or an AI
agent asks a question,

27
00:01:55.351 --> 00:01:58.272
like “How do I reset my VPN password?”,

28
00:01:58.272 --> 00:02:01.017
ServiceNow turns the query
into an embedding,

29
00:02:01.017 --> 00:02:03.376
and compares it with the stored vectors.

30
00:02:04.213 --> 00:02:08.176
The RAG Search API handles
this behind the scenes.

31
00:02:08.533 --> 00:02:12.147
It goes through the search
profile to the semantic index,

32
00:02:12.147 --> 00:02:14.511
and returns the most relevant results,

33
00:02:14.511 --> 00:02:18.079
based on similarity scores
and metadata filters.

34
00:02:18.436 --> 00:02:23.160
Those results go to the LLM,
which writes the response.

35
00:02:23.997 --> 00:02:27.280
The result is not just a keyword match.

36
00:02:27.637 --> 00:02:31.061
It's an answer grounded
in your own platform data,

37
00:02:31.061 --> 00:02:32.944
which is the essence of RAG.

38
00:02:34.164 --> 00:02:38.874
Where things get really interesting
is when you start customising this.

39
00:02:39.711 --> 00:02:43.012
RAG in ServiceNow works
as a layered pipeline,

40
00:02:43.012 --> 00:02:46.595
and each layer is a place
to add intelligence or control.

41
00:02:46.951 --> 00:02:52.318
The agent orchestration layer decides
when to call RAG, which profiles to use,

42
00:02:52.318 --> 00:02:54.392
and how to combine the sources.

43
00:02:54.749 --> 00:02:57.589
The retrieval layer
handles chunking, embedding

44
00:02:57.589 --> 00:03:02.978
and scoring, so you can change chunk sizes
or embedding models for different content.

45
00:03:03.335 --> 00:03:05.889
The ranking layer refines the results,

46
00:03:05.889 --> 00:03:11.007
for example boosting recent content,
or weighting technical documentation.

47
00:03:11.364 --> 00:03:15.693
The prompt composition layer
builds the message to the LLM.

48
00:03:16.050 --> 00:03:18.767
And the post-processing layer verifies,

49
00:03:18.767 --> 00:03:22.743
formats and cites the answer,
before the user sees it.

50
00:03:23.963 --> 00:03:29.284
This turns RAG into a programmable
system, not a fixed capability.

51
00:03:30.121 --> 00:03:31.752
For example, a scripted

52
00:03:31.752 --> 00:03:36.955
REST API could query specific search
profiles through the RAG Search API,

53
00:03:36.955 --> 00:03:42.032
re-rank the results with your own logic,
and inject them into a Now LLM prompt.

54
00:03:42.869 --> 00:03:44.187
That's where ServiceNow's

55
00:03:44.187 --> 00:03:49.133
native RAG becomes part of a broader
agentic workflow, where agents reason,

56
00:03:49.133 --> 00:03:53.163
retrieve and act, all within
the boundaries of the same platform.

57
00:03:53.520 --> 00:03:57.618
You move from asking questions
to operating workflows.

58
00:03:58.838 --> 00:04:02.570
RAG in ServiceNow isn't
a mystery box any more.

59
00:04:02.926 --> 00:04:07.962
It's a structured system that you
can tune, extend and experiment with.

60
00:04:08.799 --> 00:04:13.387
The systems that succeed won't
be the ones with the flashiest models.

61
00:04:13.744 --> 00:04:17.983
They'll be the ones that retrieve
the right context, at the right time,

62
00:04:17.983 --> 00:04:19.541
for the right task.

63
00:04:19.897 --> 00:04:24.558
Because in the end, the intelligence
of an agent depends on what it knows,

64
00:04:24.558 --> 00:04:26.931
and that depends on what it can find.
