Skip to content

Commit fbaa0b1

Browse files
ViktarStarastsenkaViktarStarastsenka
committed
Changing the AI tutorial
1 parent df0e45a commit fbaa0b1

File tree

1 file changed

+120
-65
lines changed

1 file changed

+120
-65
lines changed

‎src/uc/ai_assistant.md

Lines changed: 120 additions & 65 deletions
Original file line numberDiff line numberDiff line change
@@ -1,94 +1,149 @@
1-
Imagine you are building a smart AI assistant that:
2-
- Remembers chats, but only temporarily
3-
- Thinks by meaning, not by matching exact words
4-
- Cleans up after itself, with no manual scripts
1+
Build a production-ready AI assistant that remembers conversations semantically and auto-cleans using Redis 8’s new features.
52

6-
Redis 8 has two powerful capabilities to make this happen:
7-
- Field-level expiration: Let individual chat messages expire on their own
8-
- Vector similarity search: Find past messages based on meaning, not keywords
3+
This smart AI assistant offers:
4+
- **Ephemeral Memory**: Messages expire automatically at different intervals
5+
- **Semantic Recall**: Search past chats by meaning, not just keywords
6+
- **Zero Maintenance**: No cleanup scripts or cron jobs needed
7+
- **Multi-User Support**: Memory isolated by user and session
8+
- **Hybrid Search**: Combine text and vector search for better results
99

10-
Let’s dive in.
10+
**Note**: [Redis 8](https://hub.docker.com/_/redis/tags) is required for this tutorial Redis 8 since it introduces `HSETEX`, which lets you set TTL per hash field, ideal for ephemeral chat memory without complex expiration logic.
1111

12-
### Short-Term Memory with Field-Level Expiry
13-
Each chat session is stored as a Redis Hash.
14-
Each message is a field in the hash.
15-
Redis 8’s new HSETEX command allows you to assign a TTL to each field, perfect for building ephemeral, session-level memory.
12+
### Hierarchical Memory Structure
13+
Store chat sessions as Redis hashes. Each message is a field with its own TTL.
1614

1715
```redis:[run_confirmation=true] Upload Session Data
18-
// session:42 is the session ID
19-
// msg:<timestamp> ensures uniqueness and traceability
20-
HSETEX session:42 msg:1717935301 120 "Hi Chatbot!"
21-
HSETEX session:42 msg:1717935361 180 "What can you do?"
22-
HSETEX session:42 msg:1717935440 90 "Can you remind me about my tasks?"
23-
HSETEX session:42 msg:1717935720 30 "What's the news today?"
16+
// User-scoped sessions for isolation
17+
// Pattern: user:{user_id}:session:{session_id}
18+
HSETEX user:alice:session:morning msg:1717935301 3600 "Good morning! What's my schedule today?"
19+
HSETEX user:alice:session:morning msg:1717935361 1800 "Remind me about the team meeting at 2 PM"
20+
HSETEX user:alice:session:morning msg:1717935420 900 "What's the weather forecast?"
21+
HSETEX user:alice:session:morning msg:1717935480 300 "Thanks, that's all for now"
22+
23+
// Different user, same session pattern
24+
HSETEX user:bob:session:work msg:1717935500 7200 "I need to prepare for the client presentation"
25+
HSETEX user:bob:session:work msg:1717935560 3600 "What are the key points I should cover?"
26+
2427
```
2528

26-
Each field automatically expires after its TTL (in seconds).
27-
No need for cron jobs or background workers.
28-
What you get:
29-
- Clean memory
30-
- Zero manual cleanup
31-
- Session-scoped retention, just like short-term memory in humans
29+
### Memory Tiers for Different Lifetimes
30+
Control how long messages last depending on their importance.
31+
32+
```redis:[run_confirmation=true] Memory Tiers Strategy
33+
// Short-term (5 minutes) - Immediate context
34+
HSETEX user:alice:session:current msg:1717935301 300 "Current conversation context"
35+
36+
// Medium-term (30 minutes) - Session memory
37+
HSETEX user:alice:session:current msg:1717935302 1800 "Important session details"
38+
39+
// Long-term (2 hours) - Cross-session context
40+
HSETEX user:alice:session:current msg:1717935303 7200 "Key user preferences and facts"
41+
```
3242

43+
### Check Current Session Memory
44+
No manual cleanup needed; expired messages vanish automatically.
3345

34-
Try it: After a few minutes, run `HGETALL session:42` and see what's left.
46+
```redis:[run_confirmation=true] Monitor Session State Over Time
47+
// After a few minutes, run this command to see what's left.
48+
HGETALL user:alice:session:morning
49+
```
3550

36-
### Vector Search for Semantic Recall
37-
Now, your assistant needs to “recall” semantically related messages, not just match by words.
38-
To do that, you’ll:
39-
- Convert messages to vector embeddings
40-
- Store them in Redis
41-
- Use Vector Search with FT.SEARCH for semantic retrieval
51+
### Vector Search Setup for Semantic Recall
52+
Create an index to store messages as vectors for semantic search.
4253

4354
```redis:[run_confirmation=true] Create a Vector Index
44-
FT.CREATE idx:memory ON HASH PREFIX 1 memory: SCHEMA
45-
message TEXT
46-
embedding VECTOR FLAT // FLAT = exact vector search
47-
6
48-
TYPE FLOAT32
49-
DIM 8 // DIM = embedding size, DIM 8 is just for demo purposes. In real use, embeddings are usually 128–1536 dimensions.
50-
DISTANCE_METRIC COSINE // COSINE = measures semantic closeness
55+
FT.CREATE idx:ai_memory
56+
ON HASH
57+
PREFIX 1 memory:
58+
SCHEMA
59+
user_id TAG SORTABLE
60+
session_id TAG SORTABLE
61+
message TEXT WEIGHT 2.0 PHONETIC dm:en
62+
context TEXT WEIGHT 1.0
63+
timestamp NUMERIC SORTABLE
64+
embedding VECTOR HNSW 6
65+
TYPE FLOAT32
66+
DIM 8 // DIM = embedding size, DIM 8 is just for demo purposes. In real use, embeddings are usually 128–1536 dimensions.
67+
DISTANCE_METRIC COSINE // COSINE = measures semantic closeness
68+
INITIAL_CAP 10000
69+
M 16
70+
EF_CONSTRUCTION 200
5171
```
5272

53-
Now, let’s add entries for your chatbots:
73+
Add sample vectorized messages (embedding dims are demo-sized):
5474

5575
```redis:[run_confirmation=true] Add entries for the chatbot
56-
// Embeddings are stored as binary FLOAT32 vectors - this is a compact format required by Redis Vector Serch indexes
57-
HSET memory:1 message "Book a dentist appointment" embedding "\x00\x00\x80?\x00\x00\x00@\x00\x00@@\x00\x00\x80@\x00\x00\x00@\x00\x00\x00@"
58-
HSET memory:2 message "Remind me to water plants" embedding "\x00\x00\x80@\x00\x00\x80@\x00\x00\x80@\x00\x00\x80?\x00\x00\x80?\x00\x00@@"
59-
HSET memory:3 message "What’s the weather like?" embedding "\x00\x00@@\x00\x00\x00@\x00\x00\x00@\x00\x00\x00@\x00\x00\x80?\x00\x00\x80?"
60-
HSET memory:4 message "Cancel my gym session" embedding "\x00\x00@@\x00\x00\x00@\x00\x00\x80?\x00\x00\x80@\x00\x00\x00@\x00\x00\x00@"
61-
HSET memory:5 message "Start a new shopping list" embedding "\x00\x00\x00@\x00\x00\x00@\x00\x00\x80?\x00\x00\x80@\x00\x00\x80?\x00\x00@@"
76+
HSET memory:alice:1 user_id "alice" session_id "morning" message "I have a dentist appointment at 3 PM today" context "healthcare scheduling appointment" timestamp 1717935301 embedding "\x00\x00\x80?\x00\x00\x00@\x00\x00@@\x00\x00\x80@\x00\x00\x00@\x00\x00\x00@"
77+
HSET memory:alice:2 user_id "alice" session_id "morning" message "Remind me to water the plants in my office" context "task reminder plants office" timestamp 1717935361 embedding "\x00\x00\x80@\x00\x00\x80@\x00\x00\x80@\x00\x00\x80?\x00\x00\x80?\x00\x00@@"
78+
HSET memory:alice:3 user_id "alice" session_id "work" message "Schedule a meeting with the engineering team" context "work scheduling meeting team" timestamp 1717935420 embedding "\x00\x00@@\x00\x00\x00@\x00\x00\x00@\x00\x00\x00@\x00\x00\x80?\x00\x00\x80?"
79+
HSET memory:bob:1 user_id "bob" session_id "work" message "I need to review the quarterly sales report" context "business analysis quarterly report" timestamp 1717935480 embedding "\x00\x00@@\x00\x00\x00@\x00\x00\x80?\x00\x00\x80@\x00\x00\x00@\x00\x00\x00@"
6280
```
6381

64-
Now your messages are vectorized and ready for search.
65-
6682
### Let Chatbot Think – Semantic Search with Vectors
67-
When a user sends a new message, convert it to an embedding and run a KNN search:
68-
69-
```redis:[run_confirmation=true] Search For Similar Messages
70-
// Returns the top 3 semantically similar messages, even if no words match directly.
71-
FT.SEARCH idx:memory "*=>[KNN 3 @embedding $vec AS score]"
72-
PARAMS 2 vec "\x00\x00@@\x00\x00\x80@\x00\x00\x00@\x00\x00\x80?\x00\x00@@\x00\x00\x00@"
73-
SORTBY score
74-
DIALECT 2
83+
When a user says something new, find all related past conversations across your entire system based on semantic meaning.
84+
85+
```redis:[run_confirmation=false] Find Top 5 Related Messages By Meaning
86+
FT.SEARCH idx:ai_memory
87+
"*=>[KNN 5 @embedding $query_vec AS vector_score]"
88+
PARAMS 2 query_vec "\x00\x00@@\x00\x00\x80@\x00\x00\x00@\x00\x00\x80?\x00\x00@@\x00\x00\x00@"
89+
RETURN 6 user_id message context vector_score timestamp
90+
SORTBY vector_score ASC
91+
DIALECT 2
7592
```
7693

7794
Now your assistant “remembers” things it’s heard before - by meaning.
7895

79-
### Real-Time Session Cleanup – Redis Handles It
80-
Want to check what's still in memory?
96+
### User-Scoped Semantic Search
97+
Your AI should only recall memories from the specific user it's talking to, not leak information between users.
98+
99+
```redis:[run_confirmation=false] Find Similar Memories For Specific User Only
100+
FT.SEARCH idx:ai_memory
101+
"(@user_id:{alice}) => [KNN 3 @embedding $query_vec AS vector_score]"
102+
PARAMS 2 query_vec "\x00\x00@@\x00\x00\x80@\x00\x00\x00@\x00\x00\x80?\x00\x00@@\x00\x00\x00@"
103+
RETURN 6 user_id message context vector_score session_id
104+
SORTBY vector_score ASC
105+
DIALECT 2
106+
```
107+
108+
### Time-Bounded Semantic Search
109+
When users ask about "recent" things, limit your search to a specific time window while still using semantic matching.
110+
111+
```redis:[run_confirmation=false] Find recent similar memories (last 24 hours)
112+
FT.SEARCH idx:ai_memory
113+
"(@timestamp:[1717849200 +inf]) => [KNN 3 @embedding $query_vec AS vector_score]"
114+
PARAMS 2 query_vec "\x00\x00@@\x00\x00\x80@\x00\x00\x00@\x00\x00\x80?\x00\x00@@\x00\x00\x00@"
115+
RETURN 6 message timestamp vector_score
116+
SORTBY timestamp DESC
117+
DIALECT 2
118+
```
119+
120+
### Session-Specific Recall
121+
When users refer to something from "earlier in our conversation," search only within the current session context.
81122

82-
```redis:[run_confirmation=false] Check Sessions
83-
HGETALL session:42
84123
```
124+
FT.SEARCH idx:ai_memory
125+
"(@user_id:{alice} @session_id:{morning}) => [KNN 10 @embedding $query_vec AS vector_score]"
126+
PARAMS 2 query_vec "\x00\x00@@\x00\x00\x80@\x00\x00\x00@\x00\x00\x80?\x00\x00@@\x00\x00\x00@"
127+
RETURN 6 message context vector_score timestamp
128+
SORTBY vector_score ASC
129+
DIALECT 2
130+
```
131+
132+
### Want to check what's still in memory?
133+
134+
Only unexpired fields remain.
85135

86-
Only the unexpired fields remain. Redis does the cleanup invisibly in the background.
87-
Your assistant has a clean, focused mind at all times.
136+
```redis:[run_confirmation=false] Check Sessions
137+
HGETALL memory:alice:1
138+
HGETALL memory:alice:2
139+
HGETALL memory:alice:3
140+
HGETALL memory:bob:1
141+
HGETALL memory:bob:2
142+
```
88143

89144
### Next Steps
90145
Now that your assistant has memory and meaning, you can:
91-
- Tie session messages to store embeddings for per-session recall
92-
- Use RAG (Retrieval-Augmented Generation) by combining Redis Vector Search with LLMs
93-
- Add per-user memory: prefix session keys with a user ID (user:42:session:...)
94-
- Introduce a fallback to persistent storage for long-term memory using Redis Flex
146+
- Combine Redis Vector Search with LLMs (RAG)
147+
- Use OpenAI or sentence-transformers for embeddings
148+
- Add fallback persistent storage with Redis Flex
149+
- Manage users with ACLs, quotas, and keyspace notifications

0 commit comments

Comments
 (0)