rag-and-memory

Name: rag-and-memory
Author: DevelopersGlobal

64stars10forks

How to install rag-and-memory?

developersglobal/ai-agent-skills/rag-and-memory

$npx -y skills add DevelopersGlobal/ai-agent-skills --skill rag-and-memory

Installs into the current project.

Summary

Patterns for Retrieval-Augmented Generation (RAG) and agent memory systems. Retrieves only relevant context, prevents context bloat, and maintains coherent state across sessions.

Files · 1

View on GitHub

SKILL.md

1	## Overview
2
3	RAG and memory systems are how AI agents work with knowledge that exceeds their context window. Done well: agents give accurate, grounded answers. Done poorly: context overflow, hallucination from stale retrieval, and performance degradation.
4
5	This skill covers the design principles and failure modes of RAG and memory architectures for production AI systems.
6
7	## When to Use
8
9	- Building any AI system that needs to access external knowledge
10	- When agent context windows are being exceeded
11	- When agents need to remember information across sessions
12	- When building Q&A, document analysis, or knowledge base systems
13
14	## Process
15
16	### Step 1: Choose the Right Memory Architecture
17
18	1. Identify what the agent needs to remember:
19	- Ephemeral: Within a single session (use in-context memory)
20	- Session-persistent: Across a user's sessions (use external key-value store)
21	- Knowledge base: Organizational or domain knowledge (use vector DB + RAG)
22	- Procedural: How to do tasks (encode in SKILL.md / system prompt)
23	2. Match the memory type to the store:
24
25	\| Memory Type \| Recommended Store \|
26	\|------------\|------------------\|
27	\| In-session facts \| Context window (summarized) \|
28	\| User preferences \| Key-value store (Redis, DynamoDB) \|
29	\| Document corpus \| Vector database (Pinecone, Weaviate, pgvector) \|
30	\| Long-term facts \| Structured DB + caching \|
31
32	Verify: Each type of information the agent needs has a defined storage mechanism.
33
34	### Step 2: Design the RAG Pipeline
35
36	3. Chunking strategy: Break documents into chunks at semantic boundaries (paragraphs, sections) — not arbitrary character counts.
37	4. Embedding model: Match the embedding model to your query type. Use the same model for indexing and retrieval.
38	5. Retrieval: Retrieve top-K most semantically similar chunks. K = 3–7 is usually optimal.
39	6. Re-ranking: After retrieval, re-rank by relevance using a cross-encoder. Top K becomes top 3–5 for the prompt.
40	7. Context injection: Inject retrieved chunks into the prompt with clear source citations.
41
42	Verify: Retrieved chunks are genuinely relevant to the query before injecting into context.
43
44	### Step 3: Prevent Context Bloat
45
46	8. Summarize, don't accumulate: For long sessions, summarize previous turns rather than appending them indefinitely.
47	9. Retrieve, don't pre-load: Only load context relevant to the current query. Don't pre-load everything.
48	10. Set context budgets: Define maximum token allocations for: system prompt, retrieved context, conversation history, user message.
49	11. Compress before injecting: Summarize long retrieved documents to extract the relevant portion only.
50
51	Verify: Total prompt length is within model limits with buffer. Retrieved context is relevant to current query.
52
53	### Step 4: Handle Retrieval Failures Gracefully
54
55	12. If retrieval returns no relevant results: say so — do not hallucinate an answer.
56	13. If retrieved documents are outdated: surface the document date to the user.
57	14. If confidence is low: present the retrieved source and let the user evaluate.
58	15. Design for "no relevant information found" as a first-class outcome.
59
60	Verify: System has defined behavior for failed/empty retrieval.
61
62	### Step 5: Measure and Optimize
63
64	16. Track retrieval quality:
65	- Precision: Are retrieved chunks relevant to the query?
66	- Recall: Are relevant chunks being retrieved at all?
67	17. Track answer quality: Use RAGAS or similar evaluation framework.
68	18. Monitor: context length per query, retrieval latency, hallucination rate.
69
70	Verify: Baseline metrics established. Retrieval precision > 80%.
71
72	## Common Rationalizations (and Rebuttals)
73
74	\| Excuse \| Rebuttal \|
75	\|--------\|----------\|
76	\| "Let's just put everything in the context" \| Context bloat degrades quality and costs money. Retrieve what's needed. \|
77	\| "The model knows this from training" \| Training knowledge is stale. Use RAG for current information. \|
78	\| "Vector search is good enough without re-ranking" \| Re-ranking improves precision significantly. It's a small cost for large quality gain. \|
79	\| "We'll fix retrieval quality later" \| Poor retrieval quality compounds into poor answer quality. Fix it now. \|
80
81	## Red Flags
82
83	- Entire document corpus pre-loaded into every prompt
84	- Retrieval returning chunks from unrelated documents
85	- No defined behavior for empty retrieval results
86	- Context window regularly at 90%+ capacity
87	- Agent answering from "training knowledge" instead of retrieved documents
88	- No source citations for retrieved information
89
90	## Verification
91
92	- [ ] Memory architecture matches the type of information needed
93	- [ ] RAG pipeline: chunk → embed → retrieve → re-rank → inject
94	-

People who liked this also viewed

6 picks

Type

rag-and-memory

64stars10forks

View on GitHub

How to install rag-and-memory?

developersglobal/ai-agent-skills/rag-and-memory

$npx -y skills add DevelopersGlobal/ai-agent-skills --skill rag-and-memory

Installs into the current project.

Summary

Patterns for Retrieval-Augmented Generation (RAG) and agent memory systems. Retrieves only relevant context, prevents context bloat, and maintains coherent state across sessions.

Files · 1

View on GitHub

SKILL.md

1	## Overview
2
3	RAG and memory systems are how AI agents work with knowledge that exceeds their context window. Done well: agents give accurate, grounded answers. Done poorly: context overflow, hallucination from stale retrieval, and performance degradation.
4
5	This skill covers the design principles and failure modes of RAG and memory architectures for production AI systems.
6
7	## When to Use
8
9	- Building any AI system that needs to access external knowledge
10	- When agent context windows are being exceeded
11	- When agents need to remember information across sessions
12	- When building Q&A, document analysis, or knowledge base systems
13
14	## Process
15
16	### Step 1: Choose the Right Memory Architecture
17
18	1. Identify what the agent needs to remember:
19	- Ephemeral: Within a single session (use in-context memory)
20	- Session-persistent: Across a user's sessions (use external key-value store)
21	- Knowledge base: Organizational or domain knowledge (use vector DB + RAG)
22	- Procedural: How to do tasks (encode in SKILL.md / system prompt)
23	2. Match the memory type to the store:
24
25	\| Memory Type \| Recommended Store \|
26	\|------------\|------------------\|
27	\| In-session facts \| Context window (summarized) \|
28	\| User preferences \| Key-value store (Redis, DynamoDB) \|
29	\| Document corpus \| Vector database (Pinecone, Weaviate, pgvector) \|
30	\| Long-term facts \| Structured DB + caching \|
31
32	Verify: Each type of information the agent needs has a defined storage mechanism.
33
34	### Step 2: Design the RAG Pipeline
35
36	3. Chunking strategy: Break documents into chunks at semantic boundaries (paragraphs, sections) — not arbitrary character counts.
37	4. Embedding model: Match the embedding model to your query type. Use the same model for indexing and retrieval.
38	5. Retrieval: Retrieve top-K most semantically similar chunks. K = 3–7 is usually optimal.
39	6. Re-ranking: After retrieval, re-rank by relevance using a cross-encoder. Top K becomes top 3–5 for the prompt.
40	7. Context injection: Inject retrieved chunks into the prompt with clear source citations.
41
42	Verify: Retrieved chunks are genuinely relevant to the query before injecting into context.
43
44	### Step 3: Prevent Context Bloat
45
46	8. Summarize, don't accumulate: For long sessions, summarize previous turns rather than appending them indefinitely.
47	9. Retrieve, don't pre-load: Only load context relevant to the current query. Don't pre-load everything.
48	10. Set context budgets: Define maximum token allocations for: system prompt, retrieved context, conversation history, user message.
49	11. Compress before injecting: Summarize long retrieved documents to extract the relevant portion only.
50
51	Verify: Total prompt length is within model limits with buffer. Retrieved context is relevant to current query.
52
53	### Step 4: Handle Retrieval Failures Gracefully
54
55	12. If retrieval returns no relevant results: say so — do not hallucinate an answer.
56	13. If retrieved documents are outdated: surface the document date to the user.
57	14. If confidence is low: present the retrieved source and let the user evaluate.
58	15. Design for "no relevant information found" as a first-class outcome.
59
60	Verify: System has defined behavior for failed/empty retrieval.
61
62	### Step 5: Measure and Optimize
63
64	16. Track retrieval quality:
65	- Precision: Are retrieved chunks relevant to the query?
66	- Recall: Are relevant chunks being retrieved at all?
67	17. Track answer quality: Use RAGAS or similar evaluation framework.
68	18. Monitor: context length per query, retrieval latency, hallucination rate.
69
70	Verify: Baseline metrics established. Retrieval precision > 80%.
71
72	## Common Rationalizations (and Rebuttals)
73
74	\| Excuse \| Rebuttal \|
75	\|--------\|----------\|
76	\| "Let's just put everything in the context" \| Context bloat degrades quality and costs money. Retrieve what's needed. \|
77	\| "The model knows this from training" \| Training knowledge is stale. Use RAG for current information. \|
78	\| "Vector search is good enough without re-ranking" \| Re-ranking improves precision significantly. It's a small cost for large quality gain. \|
79	\| "We'll fix retrieval quality later" \| Poor retrieval quality compounds into poor answer quality. Fix it now. \|
80
81	## Red Flags
82
83	- Entire document corpus pre-loaded into every prompt
84	- Retrieval returning chunks from unrelated documents
85	- No defined behavior for empty retrieval results
86	- Context window regularly at 90%+ capacity
87	- Agent answering from "training knowledge" instead of retrieved documents
88	- No source citations for retrieved information
89
90	## Verification
91
92	- [ ] Memory architecture matches the type of information needed
93	- [ ] RAG pipeline: chunk → embed → retrieve → re-rank → inject
94	-

People who liked this also viewed

6 picks

Type