aprender-rag 0.63.0

Pure-Rust Retrieval-Augmented Generation pipeline built on Trueno
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135

[package]
name = "aprender-rag"
version.workspace = true
edition = "2021"
authors = ["Pragmatic AI Labs"]
description = "Pure-Rust Retrieval-Augmented Generation pipeline built on Trueno"
license = "MIT"
repository = "https://github.com/paiml/trueno-rag"
keywords = ["rag", "retrieval", "embeddings", "nlp", "vector-search"]
categories = ["text-processing", "science"]
readme = "README.md"
rust-version = "1.75"
exclude = [
    "target/",
    "book/",
    "docs/",
    "examples/",
    "tests/",
    ".pmat/",
    ".pmat-work/",
    ".pmat-tickets/",
    ".github/",
    ".vscode/",
    ".idea/",
    "*.profraw",
    "*.profdata",
    "proptest-regressions/",
]

[lib]
name = "aprender_rag"

[dependencies]
# Core compute
trueno = { workspace = true }
trueno-db = { workspace = true }

# Async runtime
tokio = { version = "1.49", features = ["full"] }

# Serialization
serde = { version = "1.0", features = ["derive"] }
serde_json = "1.0"

# Error handling
thiserror = "2.0"

# Utilities
uuid = { version = "1.11", features = ["v4", "serde"] }
async-trait = "0.1"

# Compression for index storage (GH-2) - via batuta-common shared module
batuta-common = { workspace = true }
lz4_flex = { version = "0.11", optional = true }
zstd = { version = "0.13", optional = true }
bincode = { version = "1.3", optional = true }  # Binary serialization for indices

# SQLite+FTS5 persistent index (sqlite-rag-integration spec)
# FTS5 is included in bundled SQLite by default (SQLITE_ENABLE_FTS5)
rusqlite = { version = "0.32", features = ["bundled"], optional = true }

# Real embeddings via fastembed (GH-1) - optional, adds ONNX Runtime dependency
fastembed = { version = "5", optional = true }

# NVIDIA Embed Nemotron 8B via realizar GGUF infrastructure (GH-3)
# Large model (~4-32GB depending on quantization), requires GGUF file
realizar = { version = "0.63.0", path = "../aprender-serve", package = "aprender-serve", default-features = false, optional = true }

# Speech-to-text via whisper-apr (Whisper ASR, .apr model format)
# Sovereign stack: whisper-apr + aprender + trueno for pure-Rust transcription
whisper-apr = { version = "0.2", optional = true, default-features = false, features = ["std", "symphonia"] }

# Eval: LLM-as-judge via Anthropic API (PMAT-015)
reqwest = { version = "0.12", features = ["json", "rustls-tls"], default-features = false, optional = true }
sha2 = { version = "0.10", optional = true }
rand = { version = "0.9", optional = true }

[build-dependencies]
serde = { version = "1", features = ["derive"] }
serde_yaml_ng = "0.10"

[dev-dependencies]
tokio-test = "0.4"
proptest = "1.9"
criterion = { version = "0.6", features = ["html_reports"] }
# HELIX-IDEA-005 Phase 2 (FALSIFY-HYBRID-004): seedable PRNG for the
# 5k-doc deterministic synthetic corpus that the build-perf gate
# ingests. ChaCha8 is the same family aprender-core uses for the
# HNSW recall fixture.
rand = "0.9"
rand_chacha = "0.9"

[[bench]]
name = "retrieval"
harness = false

[features]
default = ["sqlite"]
# SQLite+FTS5 persistent BM25 index (sqlite-rag-integration spec)
sqlite = ["dep:rusqlite"]
# Compressed index serialization (GH-2)
compression = ["dep:lz4_flex", "dep:zstd", "dep:bincode"]
# Real semantic embeddings via fastembed (GH-1)
# Adds ONNX Runtime dependency for production-quality embeddings
embeddings = ["dep:fastembed"]
# NVIDIA Embed Nemotron 8B via realizar GGUF infrastructure (GH-3)
# Llama 3.1 8B-based embedding model with 4096-dim output
nemotron = ["dep:realizar"]
# ColBERT-style multi-vector retrieval with WARP algorithm (PMAT-003)
# Enables token-level late interaction scoring with residual quantization
multivector = []
# Speech-to-text via whisper-apr (Whisper ASR engine)
# Sovereign stack: whisper-apr + aprender + trueno for pure-Rust transcription
transcription = ["dep:whisper-apr"]
# LLM-as-judge evaluation via Anthropic API (PMAT-015)
eval = ["dep:reqwest", "dep:sha2", "dep:rand"]
# Image OCR via tesseract CLI (PMAT-023)
# Requires tesseract-ocr system package installed on $PATH
ocr = []

[lints]
workspace = true

[package.metadata.docs.rs]
all-features = true
rustdoc-args = ["--generate-link-to-definition"]

[package.metadata.release]
shared-version = true

[[package.metadata.release.pre-release-replacements]]
file = "CHANGELOG.md"
search = "## \\[Unreleased\\]"
replace = "## [{{version}}] - {{date}}"