You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
**InvokeModel:** Uses **model-specific JSON** in the request body. The expected schema varies by model family.
41
+
40
42
**Converse API:** Use when you want a **consistent message-based interface across Bedrock models** with optional inference settings, tool config, guardrails, and prompt variables.
41
43
44
+
**Inference settings:**
45
+
Low **temperature** = more deterministic and consistent
46
+
Higher **temperature** = more creative and variable
47
+
Lower **topP** = tighter token selection
48
+
42
49
---
43
50
44
51
## Agents vs. Knowledge Bases
@@ -65,6 +72,7 @@ Review this page 5 minutes before the exam.
65
72
- Guardrails apply to **both inputs AND outputs**
66
73
- Must be explicitly applied per API call via `guardrailIdentifier` + `guardrailVersion`
67
74
- PII modes: **Redact** (mask with placeholder) vs. **Block** (reject request/response)
75
+
-**Denied Topics** = business-policy blocking at the topic level, not just exact keyword blocking
68
76
69
77
---
70
78
@@ -76,6 +84,14 @@ Review this page 5 minutes before the exam.
- The same prompt run multiple times will usually produce **nearly identical output**
115
+
116
+
**High temperature**
117
+
- Increases variation and creativity
118
+
- Best for: brainstorming, marketing copy, ideation, and multiple stylistic variants
119
+
120
+
**TopP**
121
+
- Narrows or widens the pool of candidate tokens before sampling
122
+
- Lower `topP` = tighter, safer token selection
123
+
- Higher `topP` = broader selection and more variety
124
+
125
+
::: tip
126
+
For the exam, the simplest rule is:
127
+
-**Need consistency and precision?** Lower **temperature**
128
+
-**Need more variety and creativity?** Raise **temperature**
129
+
:::
130
+
131
+
::: warning Common Confusion
132
+
Low temperature improves **consistency**, but it does **not guarantee correctness**. A deterministic answer can still be wrong if the prompt, context, or retrieval is wrong.
133
+
:::
134
+
80
135
---
81
136
82
137
## 1.3 Data Management & RAG Pipelines
@@ -140,6 +195,48 @@ Embeddings are numerical vector representations of text. Semantically similar te
140
195
|**Titan Text Embeddings v2**| AWS | General purpose, AWS-native RAG (configurable dimensions) |
**Embedding dimensionality** is the number of values in each vector. It is an architecture tradeoff between **semantic richness** and **efficiency**.
201
+
202
+
| Choice | Advantage | Cost / Trade-off | Best Fit |
203
+
|---|---|---|---|
204
+
|**Higher dimensions**| Captures more subtle semantic meaning and nuance | More storage, more memory use, more compute during similarity search | Complex semantic retrieval, nuanced documents, harder query understanding |
Poor-quality source data leads to poor embeddings, poor retrieval, and poor model answers. If the scenario is about **validating structured data before it enters a GenAI pipeline**, think about upstream data-quality controls rather than guardrails or prompt changes.
222
+
223
+
**AWS Glue Data Quality** is used to:
224
+
- Define explicit quality rules for pipeline data
225
+
- Evaluate records against those rules before downstream processing
226
+
- Produce quality scores and flag failures
227
+
- Stop or quarantine bad data before it reaches embeddings, vector stores, or foundation models
228
+
229
+
**Examples of rules:**
230
+
-`patient_id` must not be null
231
+
-`admission_date` must match a valid ISO date format
232
+
- Required columns must be present
233
+
- Value ranges and uniqueness constraints must hold
234
+
235
+
::: tip
236
+
Use **AWS Glue Data Quality** when the problem is **"validate and block poor-quality structured data in an ETL pipeline before model use."**
237
+
Use **Guardrails** when the problem is filtering model inputs/outputs at inference time.
238
+
:::
239
+
143
240
---
144
241
145
242
## 1.4 Vector Stores
@@ -163,6 +260,52 @@ Embeddings are numerical vector representations of text. Semantically similar te
163
260
- Integrated with Bedrock Knowledge Bases via IAM service-linked role
164
261
- Does NOT support standard OpenSearch full-text features (custom analyzers, etc.)
165
262
263
+
### OpenSearch as a Vector Search Engine
264
+
265
+
Amazon OpenSearch is one of the standard AWS answers when the scenario requires **storing embeddings** and retrieving **semantically similar items** with low latency.
266
+
267
+
**How to think about it:**
268
+
- Store embedding vectors in the index
269
+
- Run **similarity search** against those vectors
270
+
- Retrieve the nearest matches for use in RAG, recommendations, or "find similar" features
271
+
- Use it when the system needs to handle **large embedding collections at scale** with fast retrieval
272
+
273
+
**Important terminology:**
274
+
-**k-nearest neighbor (k-NN)** = retrieve the `k` most similar vectors to the query vector
275
+
-**Vector search / similarity search** = search by semantic closeness, not exact keyword match
276
+
-**OpenSearch Serverless Vector Engine** = the Bedrock-friendly serverless option
277
+
-**OpenSearch Service with the k-NN plugin** = the broader managed-service variant
278
+
279
+
For exam purposes, if the prompt says **store embeddings + low-latency semantic similarity search**, OpenSearch is a strong answer. If it specifically mentions **Bedrock Knowledge Bases**, default to **OpenSearch Serverless**.
280
+
281
+
::: tip
282
+
OpenSearch is a strong answer for **semantic search**, **content recommendation**, and other "find similar items" patterns because it combines **vector similarity search** with optional **hybrid keyword search** in a managed AWS service.
283
+
:::
284
+
285
+
### Retrieval Vocabulary You Should Know
286
+
287
+
| Concept | Meaning | Why It Matters |
288
+
|---|---|---|
289
+
|**k-NN**| Return the `k` nearest vectors to a query vector | Core concept behind semantic retrieval |
290
+
|**ANN (Approximate Nearest Neighbor)**| Faster nearest-neighbor search that trades a bit of exactness for speed | Real systems use ANN for low-latency search at scale |
291
+
|**HNSW**| A graph-based ANN indexing approach | Often appears when discussing high-performance vector search internals |
292
+
|**Similarity Search**| Search by vector closeness instead of exact keyword match | Core retrieval behavior in RAG |
293
+
|**Semantic Search**| Search by meaning, not literal wording | Explains why embeddings are useful |
294
+
|**Hybrid Search**| Combine semantic search with keyword search | Often improves recall on enterprise documents |
|**Reranking**| Re-score an initial result set with a stronger model or scoring step | Improves relevance after the first retrieval pass |
297
+
298
+
### Retrieval Quality Trade-Offs
299
+
300
+
-**Recall** = how many truly relevant chunks are retrieved
301
+
-**Precision** = how many retrieved chunks are actually relevant
302
+
-**Higher recall** usually means retrieving more candidates, which can increase cost and noise
303
+
-**Higher precision** usually means tighter filtering or reranking, which can improve answer quality
304
+
305
+
::: tip
306
+
When the exam describes a retrieval system that misses relevant information, think about **recall**. When it retrieves too much irrelevant context, think about **precision**, metadata filtering, or reranking.
307
+
:::
308
+
166
309
---
167
310
168
311
## 1.5 Compliance, Data Residency & Security
@@ -182,6 +325,29 @@ Embeddings are numerical vector representations of text. Semantically similar te
182
325
Amazon Bedrock does **not** use customer prompts, completions, or training data to train the underlying foundation models. This is a built-in data privacy guarantee — any answer suggesting otherwise is wrong.
183
326
:::
184
327
328
+
### AWS KMS and Compliance Scenarios
329
+
330
+
AWS KMS is the core encryption-at-rest control when the scenario emphasizes **customer control over keys**, auditability, or regulated workloads.
331
+
332
+
**What customer-managed KMS keys give you:**
333
+
- Control over key policies and who can use the key
334
+
- Control over rotation policy and lifecycle management
335
+
- The ability to disable or revoke key usage if needed
336
+
- Auditable key usage through AWS logging
337
+
338
+
**Common integrations to know:**
339
+
-**Amazon S3** for document storage and RAG source data
340
+
-**Amazon DynamoDB** for application state, user data, or agent metadata
341
+
-**Amazon EBS** for encrypted block storage attached to EC2 workloads
342
+
343
+
**Audit trail:**
344
+
-**AWS CloudTrail** logs KMS API activity and key usage events
345
+
- This is the service to choose when the question asks for compliance-oriented auditing or evidence of encryption-key usage
346
+
347
+
**Exam framing:**
348
+
- If the scenario says the organization needs **full control over encryption keys**, prefer **customer-managed KMS keys**
349
+
- If the scenario also stresses **HIPAA**, auditability, and strict security controls, customer-managed KMS keys plus CloudTrail is a strong signal
350
+
185
351
---
186
352
187
353
## 1.6 Model Customization: Bedrock Fine-Tuning vs. SageMaker vs. RAG
0 commit comments