llm-app-patterns

Name: llm-app-patterns
Author: tranhieutt

69stars41forks

How to install llm-app-patterns?

tranhieutt/software_development_department/llm-app-patterns

$npx -y skills add tranhieutt/software_development_department --skill llm-app-patterns

Installs into the current project.

Summary

Provides architectural patterns for LLM-powered applications and AI assistants, including prompt engineering, RAG, agent loops, conversation management, and evaluation. Use when building AI-based features, chatbots, or complex AI system architectures.

Files · 1

View on GitHub

SKILL.md

1	# LLM Application & AI Assistant Patterns
2
3	## Resources
4
5
6
7	## Architecture decision matrix
8
9	\| Pattern \| Use when \| Cost \|
10	\|---\|---\|---\|
11	\| Simple RAG \| FAQ, docs Q&A \| Low \|
12	\| Hybrid RAG (semantic + BM25) \| Mixed query types \| Medium \|
13	\| Function calling \| Structured tool use \| Low \|
14	\| ReAct agent \| Multi-step reasoning \| Medium \|
15	\| Plan-and-execute \| Complex decomposable tasks \| High \|
16	\| Multi-agent \| Research, critique-refine \| Very High \|
17
18	## RAG: critical config numbers
19
20	```python
21	CHUNK_CONFIG = {
22	"chunk_size": 512, # tokens — sweet spot for most docs
23	"chunk_overlap": 50, # prevents context loss at boundaries
24	"separators": ["\n\n", "\n", ". ", " "],
25	}
26	# Hybrid search alpha: 1.0=semantic only, 0.0=BM25 only, 0.5=balanced
27	```
28
29	## RAG: retrieval strategies
30
31	```python
32	# Basic: semantic search
33	results = vector_db.similarity_search(embed(query), top_k=5)
34
35	# Better: hybrid (semantic + keyword via RRF)
36	def hybrid_search(query, alpha=0.5):
37	return rrf_merge(vector_db.search(query), bm25_search(query), alpha)
38
39	# Best for recall: multi-query (3 variations, deduplicate)
40	queries = llm.generate_variations(query, n=3)
41	results = deduplicate([semantic_search(q) for q in queries])
42	```
43
44	## RAG: generation prompt template
45
46	```python
47	RAG_PROMPT = """Answer based ONLY on the context below.
48	If insufficient, say "I don't have enough information."
49
50	Context: {context}
51	Question: {question}
52	Answer:"""
53	```
54
55	## Agent: function calling loop
56
57	```python
58	messages = [{"role": "user", "content": question}]
59	while True:
60	response = llm.chat(messages=messages, tools=TOOLS, tool_choice="auto")
61	if not response.tool_calls:
62	return response.content
63	for call in response.tool_calls:
64	result = execute_tool(call.name, call.arguments)
65	messages.append({"role": "tool", "tool_call_id": call.id, "content": str(result)})
66	```
67
68	## Production: caching (only temperature=0 responses)
69
70	```python
71	def get_or_generate(prompt, model, **kwargs):
72	deterministic = kwargs.get("temperature", 1.0) == 0
73	if deterministic:
74	key = sha256(f"{model}:{prompt}:{json.dumps(kwargs, sort_keys=True)}")
75	if cached := redis.get(key): return cached
76	response = llm.generate(prompt, model=model, **kwargs)
77	if deterministic: redis.setex(key, 3600, response)
78	return response
79	```
80
81	## Production: retry + fallback
82
83	```python
84	from tenacity import retry, wait_exponential, stop_after_attempt
85
86	@retry(wait=wait_exponential(multiplier=1, min=4, max=60), stop=stop_after_attempt(5))
87	def call_llm(prompt): return llm.generate(prompt)
88
89	# Fallback chain
90	for model in [primary] + fallbacks:
91	try: return llm.generate(prompt, model=model)
92	except (RateLimitError, APIError): continue
93	```
94
95	## LLMOps: key metrics
96
97	```
98	Latency : p50, p99 response time
99	Quality : satisfaction (thumbs), task completion %, hallucination rate
100	Cost : cost_per_request, tokens_per_request, cache_hit_rate
101	Health : error_rate, timeout_rate, retry_rate
102	```
103
104	## Embedding model selection
105
106	\| Model \| Dims \| Cost \| Use \|
107	\|---\|---\|---\|---\|
108	\| text-embedding-3-small \| 1536 \| $0.02/1M \| Most cases \|
109	\| text-embedding-3-large \| 3072 \| $0.13/1M \| High accuracy \|
110	\| bge-large (local) \| 1024 \| Free \| Self-hosted \|

People who liked this also viewed

6 picks

Type

llm-app-patterns

69stars41forks

View on GitHub

How to install llm-app-patterns?

tranhieutt/software_development_department/llm-app-patterns

$npx -y skills add tranhieutt/software_development_department --skill llm-app-patterns

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# LLM Application & AI Assistant Patterns
2
3	## Resources
4
5
6
7	## Architecture decision matrix
8
9	\| Pattern \| Use when \| Cost \|
10	\|---\|---\|---\|
11	\| Simple RAG \| FAQ, docs Q&A \| Low \|
12	\| Hybrid RAG (semantic + BM25) \| Mixed query types \| Medium \|
13	\| Function calling \| Structured tool use \| Low \|
14	\| ReAct agent \| Multi-step reasoning \| Medium \|
15	\| Plan-and-execute \| Complex decomposable tasks \| High \|
16	\| Multi-agent \| Research, critique-refine \| Very High \|
17
18	## RAG: critical config numbers
19
20	```python
21	CHUNK_CONFIG = {
22	"chunk_size": 512, # tokens — sweet spot for most docs
23	"chunk_overlap": 50, # prevents context loss at boundaries
24	"separators": ["\n\n", "\n", ". ", " "],
25	}
26	# Hybrid search alpha: 1.0=semantic only, 0.0=BM25 only, 0.5=balanced
27	```
28
29	## RAG: retrieval strategies
30
31	```python
32	# Basic: semantic search
33	results = vector_db.similarity_search(embed(query), top_k=5)
34
35	# Better: hybrid (semantic + keyword via RRF)
36	def hybrid_search(query, alpha=0.5):
37	return rrf_merge(vector_db.search(query), bm25_search(query), alpha)
38
39	# Best for recall: multi-query (3 variations, deduplicate)
40	queries = llm.generate_variations(query, n=3)
41	results = deduplicate([semantic_search(q) for q in queries])
42	```
43
44	## RAG: generation prompt template
45
46	```python
47	RAG_PROMPT = """Answer based ONLY on the context below.
48	If insufficient, say "I don't have enough information."
49
50	Context: {context}
51	Question: {question}
52	Answer:"""
53	```
54
55	## Agent: function calling loop
56
57	```python
58	messages = [{"role": "user", "content": question}]
59	while True:
60	response = llm.chat(messages=messages, tools=TOOLS, tool_choice="auto")
61	if not response.tool_calls:
62	return response.content
63	for call in response.tool_calls:
64	result = execute_tool(call.name, call.arguments)
65	messages.append({"role": "tool", "tool_call_id": call.id, "content": str(result)})
66	```
67
68	## Production: caching (only temperature=0 responses)
69
70	```python
71	def get_or_generate(prompt, model, **kwargs):
72	deterministic = kwargs.get("temperature", 1.0) == 0
73	if deterministic:
74	key = sha256(f"{model}:{prompt}:{json.dumps(kwargs, sort_keys=True)}")
75	if cached := redis.get(key): return cached
76	response = llm.generate(prompt, model=model, **kwargs)
77	if deterministic: redis.setex(key, 3600, response)
78	return response
79	```
80
81	## Production: retry + fallback
82
83	```python
84	from tenacity import retry, wait_exponential, stop_after_attempt
85
86	@retry(wait=wait_exponential(multiplier=1, min=4, max=60), stop=stop_after_attempt(5))
87	def call_llm(prompt): return llm.generate(prompt)
88
89	# Fallback chain
90	for model in [primary] + fallbacks:
91	try: return llm.generate(prompt, model=model)
92	except (RateLimitError, APIError): continue
93	```
94
95	## LLMOps: key metrics
96
97	```
98	Latency : p50, p99 response time
99	Quality : satisfaction (thumbs), task completion %, hallucination rate
100	Cost : cost_per_request, tokens_per_request, cache_hit_rate
101	Health : error_rate, timeout_rate, retry_rate
102	```
103
104	## Embedding model selection
105
106	\| Model \| Dims \| Cost \| Use \|
107	\|---\|---\|---\|---\|
108	\| text-embedding-3-small \| 1536 \| $0.02/1M \| Most cases \|
109	\| text-embedding-3-large \| 3072 \| $0.13/1M \| High accuracy \|
110	\| bge-large (local) \| 1024 \| Free \| Self-hosted \|

People who liked this also viewed

6 picks

Type