faiss

Name: faiss
Author: AlexAI-MCP

132stars24forks

How to install faiss?

alexai-mcp/hermes-ccc/faiss

$npx -y skills add AlexAI-MCP/hermes-CCC --skill faiss

Installs into the current project.

Summary

Facebook AI Similarity Search — ultra-fast vector similarity search for large-scale local embedding retrieval without a database server.

Files · 1

View on GitHub

SKILL.md

1	# FAISS — Fast Vector Similarity Search
2
3	FAISS (Facebook AI Similarity Search) is a library for efficient similarity search over large collections of vectors. No server required — runs entirely in-process.
4
5	## When to Use FAISS
6
7	- Millions of vectors, need maximum speed
8	- Can't run a server (embedded use case)
9	- Research / prototyping at scale
10	- Custom index types (HNSW, IVF, PQ)
11
12	## Setup
13
14	```bash
15	pip install faiss-cpu # CPU only
16	pip install faiss-gpu # GPU (requires CUDA)
17	pip install sentence-transformers numpy
18	```
19
20	---
21
22	## Basic Flat Index (Exact Search)
23
24	```python
25	import faiss
26	import numpy as np
27	from sentence_transformers import SentenceTransformer
28
29	model = SentenceTransformer("all-MiniLM-L6-v2") # dim=384
30	DIM = 384
31
32	# Build index
33	documents = [
34	"Python async programming",
35	"Machine learning with PyTorch",
36	"Docker container management",
37	"GraphQL API design",
38	]
39
40	embeddings = model.encode(documents).astype(np.float32)
41
42	# Flat L2 index (exact, brute force)
43	index = faiss.IndexFlatL2(DIM)
44	index.add(embeddings)
45	print(f"Index size: {index.ntotal}")
46
47	# Search
48	query = "how to do async in Python?"
49	q_vec = model.encode([query]).astype(np.float32)
50
51	distances, indices = index.search(q_vec, k=3)
52	for i, idx in enumerate(indices[0]):
53	print(f"[{distances[0][i]:.3f}] {documents[idx]}")
54	```
55
56	---
57
58	## Cosine Similarity (Normalize first)
59
60	```python
61	# For cosine similarity: normalize vectors, use IndexFlatIP (inner product)
62	def normalize(vecs):
63	norms = np.linalg.norm(vecs, axis=1, keepdims=True)
64	return vecs / np.maximum(norms, 1e-8)
65
66	embeddings_norm = normalize(embeddings.astype(np.float32))
67	index_cos = faiss.IndexFlatIP(DIM)
68	index_cos.add(embeddings_norm)
69
70	q_norm = normalize(model.encode([query]).astype(np.float32))
71	scores, indices = index_cos.search(q_norm, k=3)
72	```
73
74	---
75
76	## IVF Index (Fast Approximate Search)
77
78	```python
79	# For large datasets (>100k vectors)
80	N_CLUSTERS = 100 # sqrt(N) is a good heuristic
81
82	quantizer = faiss.IndexFlatL2(DIM)
83	index_ivf = faiss.IndexIVFFlat(quantizer, DIM, N_CLUSTERS)
84
85	# Must train before adding
86	index_ivf.train(embeddings)
87	index_ivf.add(embeddings)
88	index_ivf.nprobe = 10 # search 10 clusters (higher = more accurate, slower)
89
90	distances, indices = index_ivf.search(q_vec, k=5)
91	```
92
93	---
94
95	## HNSW Index (Best Speed/Accuracy Tradeoff)
96
97	```python
98	index_hnsw = faiss.IndexHNSWFlat(DIM, 32) # 32 = M parameter
99	index_hnsw.add(embeddings)
100
101	distances, indices = index_hnsw.search(q_vec, k=5)
102	```
103
104	---
105
106	## Save and Load Index
107
108	```python
109	# Save
110	faiss.write_index(index, "my_index.faiss")
111
112	# Load
113	index = faiss.read_index("my_index.faiss")
114	```
115
116	---
117
118	## With Metadata (store separately)
119
120	```python
121	import pickle
122
123	# FAISS only stores vectors — store metadata separately
124	metadata = [{"id": i, "text": doc} for i, doc in enumerate(documents)]
125
126	# Save both
127	faiss.write_index(index, "vectors.faiss")
128	with open("metadata.pkl", "wb") as f:
129	pickle.dump(metadata, f)
130
131	# Load and query
132	index = faiss.read_index("vectors.faiss")
133	with open("metadata.pkl", "rb") as f:
134	metadata = pickle.load(f)
135
136	distances, indices = index.search(q_vec, k=3)
137	for idx in indices[0]:
138	print(metadata[idx]["text"])
139	```
140
141	---
142
143	## GPU Acceleration
144
145	```python
146	res = faiss.StandardGpuResources()
147	index_gpu = faiss.index_cpu_to_gpu(res, 0, index) # GPU 0
148	distances, indices = index_gpu.search(q_vec, k=5)
149	```
150
151	---
152
153	## Index Selection Guide
154
155	\| Index \| Size \| Speed \| Accuracy \| Use Case \|
156	\|-------\|------\|-------\|----------\|----------\|
157	\| IndexFlatL2 \| Any \| Slow \| 100% \| <100k vectors, exact needed \|
158	\| IndexFlatIP \| Any \| Slow \| 100% \| Cosine similarity \|
159	\| IndexIVFFlat \| Large \| Fast \| ~95% \| 100k–10M vectors \|
160	\| IndexHNSWFlat \| Medium \| Very fast \| ~99% \| Best general choice \|
161	\| IndexIVFPQ \| Huge \| Fastest \| ~90% \| Billions, memory-constrained \|

People who liked this also viewed

6 picks

Type

web-design-guidelinesReview UI code for Web Interface Guidelines compliance. Use when asked to "review my UI", "check accessibility", "audit design", "review UX", or "check my site against best practices".SkillsJul 2026483k29k
lark-slides飞书幻灯片：创建和编辑幻灯片。创建演示文稿、读取幻灯片内容、管理幻灯片页面（创建、删除、读取、局部替换）。当用户需要创建或编辑幻灯片、读取或修改单个页面时使用。当用户给出 doubao.com 的 /slides/ URL/token 时，也应直接使用本 skill，不要因为域名不是飞书而回退到 WebFetch；路由依据是 URL 路径模式和 token，而不是域名。不负责：云文档内容编辑（走 SkillsJul 2026332k16k
design-taste-frontendTaste-Skill - gives your AI good taste. stops the AI from generating boring, generic slop SkillsJul 2026280k66k
ui-ux-pro-maxUI/UX design intelligence for web and mobile. Searchable local database with 84 styles, 192 color palettes, 74 font pairings, 192 product types, 98 UX guidelines, 104 icon entries, 16 GSAP motion presSkillsJul 2026280k109k
caveman-helpQuick-reference card for all caveman modes, skills, and commands. One-shot display, not a persistent mode. Trigger: /caveman-help, "caveman help", "what caveman commands", "how do I use caveman".SkillsJul 2026249k92k
codebase-designShared vocabulary for designing deep modules. Use when the user wants to design or improve a module's interface, find deepening opportunities, decide where a seam goes, make code more testable or AI-nSkillsJul 2026225k182k

faiss

132stars24forks

View on GitHub

How to install faiss?

alexai-mcp/hermes-ccc/faiss

$npx -y skills add AlexAI-MCP/hermes-CCC --skill faiss

Installs into the current project.

Summary

Facebook AI Similarity Search — ultra-fast vector similarity search for large-scale local embedding retrieval without a database server.

Files · 1

View on GitHub

SKILL.md

1	# FAISS — Fast Vector Similarity Search
2
3	FAISS (Facebook AI Similarity Search) is a library for efficient similarity search over large collections of vectors. No server required — runs entirely in-process.
4
5	## When to Use FAISS
6
7	- Millions of vectors, need maximum speed
8	- Can't run a server (embedded use case)
9	- Research / prototyping at scale
10	- Custom index types (HNSW, IVF, PQ)
11
12	## Setup
13
14	```bash
15	pip install faiss-cpu # CPU only
16	pip install faiss-gpu # GPU (requires CUDA)
17	pip install sentence-transformers numpy
18	```
19
20	---
21
22	## Basic Flat Index (Exact Search)
23
24	```python
25	import faiss
26	import numpy as np
27	from sentence_transformers import SentenceTransformer
28
29	model = SentenceTransformer("all-MiniLM-L6-v2") # dim=384
30	DIM = 384
31
32	# Build index
33	documents = [
34	"Python async programming",
35	"Machine learning with PyTorch",
36	"Docker container management",
37	"GraphQL API design",
38	]
39
40	embeddings = model.encode(documents).astype(np.float32)
41
42	# Flat L2 index (exact, brute force)
43	index = faiss.IndexFlatL2(DIM)
44	index.add(embeddings)
45	print(f"Index size: {index.ntotal}")
46
47	# Search
48	query = "how to do async in Python?"
49	q_vec = model.encode([query]).astype(np.float32)
50
51	distances, indices = index.search(q_vec, k=3)
52	for i, idx in enumerate(indices[0]):
53	print(f"[{distances[0][i]:.3f}] {documents[idx]}")
54	```
55
56	---
57
58	## Cosine Similarity (Normalize first)
59
60	```python
61	# For cosine similarity: normalize vectors, use IndexFlatIP (inner product)
62	def normalize(vecs):
63	norms = np.linalg.norm(vecs, axis=1, keepdims=True)
64	return vecs / np.maximum(norms, 1e-8)
65
66	embeddings_norm = normalize(embeddings.astype(np.float32))
67	index_cos = faiss.IndexFlatIP(DIM)
68	index_cos.add(embeddings_norm)
69
70	q_norm = normalize(model.encode([query]).astype(np.float32))
71	scores, indices = index_cos.search(q_norm, k=3)
72	```
73
74	---
75
76	## IVF Index (Fast Approximate Search)
77
78	```python
79	# For large datasets (>100k vectors)
80	N_CLUSTERS = 100 # sqrt(N) is a good heuristic
81
82	quantizer = faiss.IndexFlatL2(DIM)
83	index_ivf = faiss.IndexIVFFlat(quantizer, DIM, N_CLUSTERS)
84
85	# Must train before adding
86	index_ivf.train(embeddings)
87	index_ivf.add(embeddings)
88	index_ivf.nprobe = 10 # search 10 clusters (higher = more accurate, slower)
89
90	distances, indices = index_ivf.search(q_vec, k=5)
91	```
92
93	---
94
95	## HNSW Index (Best Speed/Accuracy Tradeoff)
96
97	```python
98	index_hnsw = faiss.IndexHNSWFlat(DIM, 32) # 32 = M parameter
99	index_hnsw.add(embeddings)
100
101	distances, indices = index_hnsw.search(q_vec, k=5)
102	```
103
104	---
105
106	## Save and Load Index
107
108	```python
109	# Save
110	faiss.write_index(index, "my_index.faiss")
111
112	# Load
113	index = faiss.read_index("my_index.faiss")
114	```
115
116	---
117
118	## With Metadata (store separately)
119
120	```python
121	import pickle
122
123	# FAISS only stores vectors — store metadata separately
124	metadata = [{"id": i, "text": doc} for i, doc in enumerate(documents)]
125
126	# Save both
127	faiss.write_index(index, "vectors.faiss")
128	with open("metadata.pkl", "wb") as f:
129	pickle.dump(metadata, f)
130
131	# Load and query
132	index = faiss.read_index("vectors.faiss")
133	with open("metadata.pkl", "rb") as f:
134	metadata = pickle.load(f)
135
136	distances, indices = index.search(q_vec, k=3)
137	for idx in indices[0]:
138	print(metadata[idx]["text"])
139	```
140
141	---
142
143	## GPU Acceleration
144
145	```python
146	res = faiss.StandardGpuResources()
147	index_gpu = faiss.index_cpu_to_gpu(res, 0, index) # GPU 0
148	distances, indices = index_gpu.search(q_vec, k=5)
149	```
150
151	---
152
153	## Index Selection Guide
154
155	\| Index \| Size \| Speed \| Accuracy \| Use Case \|
156	\|-------\|------\|-------\|----------\|----------\|
157	\| IndexFlatL2 \| Any \| Slow \| 100% \| <100k vectors, exact needed \|
158	\| IndexFlatIP \| Any \| Slow \| 100% \| Cosine similarity \|
159	\| IndexIVFFlat \| Large \| Fast \| ~95% \| 100k–10M vectors \|
160	\| IndexHNSWFlat \| Medium \| Very fast \| ~99% \| Best general choice \|
161	\| IndexIVFPQ \| Huge \| Fastest \| ~90% \| Billions, memory-constrained \|

People who liked this also viewed

6 picks

Type