attune-rag

Lightweight, LLM-agnostic RAG pipeline with pluggable corpora. Works with Claude, Gemini, or any LLM.

• No LLM SDK at install time. All provider deps are optional extras. Two required runtime deps: structlog, jinja2.
• Pluggable corpus. Use attune-help (the default), any markdown directory, or your own CorpusProtocol.
• Returns a prompt string + citation records by default — pipeline.run() never opens a network connection. You call your own LLM however you like. Optional provider adapters ship convenience wrappers.
• Optional hybrid retrieval. QueryExpander and LLMReranker layer Claude Haiku on top of keyword retrieval to improve recall and precision — both opt-in, both fail-safe.

Why attune-rag

Most RAG libraries ship features. attune-rag ships measured quality numbers and gates merges against them. The CI badge isn't "tests pass" — it's P@1 ≥ 0.95, R@3 = 1.00, mean faithfulness ≥ 0.9686 (locked at docs/specs/release-quality-baseline/baseline-1.md) plus per-axis CPU + wall-clock perf thresholds (locked at docs/specs/downstream-validation/perf-baseline.md).

A PR that drops mean_faithfulness below 0.9686 fails CI automatically. Same for any latency hot-path regressing past mean + 2σ. That's the differentiator.

vs LangChain / LlamaIndex

	attune-rag	LangChain	LlamaIndex
Required runtime deps	2	many (transitively, ~30+)	many (~25+)
LLM SDK at install	none	bundled	bundled
Published quality regression thresholds	yes (P@1, R@3, faithfulness)	no	no
Published perf thresholds (wall + CPU)	yes	no	no
Citation primitives built-in	yes	add-on	add-on
"Get a string back, call your own LLM"	default	possible w/ effort	possible w/ effort

LangChain and LlamaIndex are fantastic frameworks if you want batteries-included orchestration. attune-rag is the alternative when you want a RAG component you can drop into an existing app without buying into a framework — and want the quality bar quantified, not implied.

Beyond drop-in retrieval, attune-rag is the grounding foundation for the attune-* family's content-quality discipline. The attune-author polish/fact-check pipeline uses attune-rag's retrieval + faithfulness primitives to verify generated help content is grounded in source material before it's marked authoritative — the same mean_faithfulness ≥ 0.9686 discipline that gates this library's own benchmarks, extended to the authoring loop.

What attune-rag is not

Honest exclusions, so you can self-disqualify if you need any of these:

• Not an agent framework. No multi-step chains, no tool-use orchestration, no agent loops.
• Not a document-parsing toolkit. Bring your markdown already-parsed; use unstructured.io or similar upstream.
• Not a vector DB integration. Keyword retrieval is the default; you wire your own vector store if you need one (an EmbeddingRetriever is on the post-freeze roadmap — see below).
• Not a one-line-install batteries-included framework. That's LangChain / LlamaIndex. attune-rag is for the case where that's too much.

Install

pip install attune-rag                     # core only
pip install 'attune-rag[attune-help]'      # + bundled help corpus
pip install 'attune-rag[claude]'           # + Claude adapter
pip install 'attune-rag[gemini]'           # + Gemini adapter
pip install 'attune-rag[all]'              # everything

Quick start — Claude

pip install 'attune-rag[attune-help,claude]'

import asyncio
from attune_rag import RagPipeline

async def main():
    pipeline = RagPipeline()  # defaults to AttuneHelpCorpus
    response, result = await pipeline.run_and_generate(
        "How do I run a security audit with attune?",
        provider="claude",
    )
    print(response)
    print("\nSources:", [h.entry.path for h in result.citation.hits])

asyncio.run(main())

Quick start — Gemini

pip install 'attune-rag[attune-help,gemini]'

response, result = await pipeline.run_and_generate(
    "...", provider="gemini", model="gemini-1.5-pro",
)

Quick start — custom corpus, any LLM

from pathlib import Path
from attune_rag import RagPipeline, DirectoryCorpus

pipeline = RagPipeline(corpus=DirectoryCorpus(Path("./my-docs")))
result = pipeline.run("How do I...?")

# Send result.augmented_prompt to whatever LLM you use.
# The pipeline itself does NOT call an LLM unless you use
# run_and_generate or call a provider adapter yourself.

📖 Building a quality corpus. See docs/USER_CORPUS_GUIDE.md for the corpus-authoring discipline that produced the bundled attune-help corpus's 100% / 100% baseline + 100% paraphrased R@3: frontmatter aliases, multi-token intent, the MIN_ALIAS_OVERLAP knob, stemmer traps, the override file pattern, and the strict-dominance measurement loop. The guide is the v0 forerunner of the v1.0.0 framework framing (user-corpus-onboarding spec).

Hybrid retrieval (optional)

QueryExpander and LLMReranker require the [claude] extra and an ANTHROPIC_API_KEY. Both are opt-in and fail-safe — any API error falls back to keyword-only order automatically.

from attune_rag import RagPipeline, LLMReranker, QueryExpander

# Reranker only (recommended for precision):
pipeline = RagPipeline(reranker=LLMReranker())

# Expander + reranker (max coverage):
pipeline = RagPipeline(
    expander=QueryExpander(),
    reranker=LLMReranker(),
)

Template editor primitives (attune_rag.editor)

Headless toolkit for tools that need to validate, lint, and refactor a template corpus — used by the attune-gui template editor and the attune-author edit CLI, but works standalone with any CorpusProtocol.

API	What it does
load_schema()	Loads template_schema.json (the v1 frontmatter contract: required type enum + name; optional tags, aliases, summary, source, hash; additionalProperties: true).
parse_frontmatter(text) / validate_frontmatter(data)	Split a template into frontmatter + body and report typed FrontmatterIssues — used by linters and editors.
lint_template(text, rel_path, corpus)	Returns Diagnostic[] for schema violations, broken [[alias]] references, and depth-marker sequence errors. 1-indexed line/col ranges.
autocomplete_tags(corpus, prefix, limit) / autocomplete_aliases(corpus, prefix, limit)	Prefix-match completions ranked by frequency (tags) or lexical proximity (aliases). Sub-ms on 1k templates.
find_references(corpus, name, kind)	Locate every alias/tag/path occurrence across body, frontmatter, and cross_links.json.
plan_rename(corpus, old, new, kind)	Build a RenamePlan (one FileEdit per affected file with unified-diff hunks) for kind="alias" or "tag". Raises RenameCollisionError on existing alias targets.
apply_rename(corpus, plan)	Atomically apply the plan (tempfile-per-file + sequential rename + drift-detection rollback). Returns the list of affected paths.

Schema, lint, and rename are pure functions over CorpusProtocol — no I/O, no global state. All three pieces are tested as a unit and used live by the attune-gui editor's /api/corpus/<id>/lint, /autocomplete, and /refactor/rename/{preview,apply} routes.

from attune_rag import DirectoryCorpus
from attune_rag.editor import lint_template, plan_rename, apply_rename

corpus = DirectoryCorpus(Path("./templates")).load()

# Validate a template before saving
diagnostics = lint_template(
    text=Path("./templates/concepts/foo.md").read_text(),
    rel_path="concepts/foo.md",
    corpus=corpus,
)

# Rename an alias across the whole corpus
plan = plan_rename(corpus, old="oldname", new="newname", kind="alias")
print(f"Affects {len(plan.edits)} files")
affected = apply_rename(corpus, plan)

Dashboard

attune-rag dashboard show    # live terminal dashboard
attune-rag dashboard render --out report.html  # HTML snapshot

Quality baselines

attune-rag locks two baselines, both gated by CI. Thresholds are empirically derived (mean ± 2σ) from back-to-back benchmark runs on an unchanged HEAD — grounded, not guessed.

Retrieval + faithfulness

Metric	Threshold (current)	Source
precision_at_1	≥ 0.95	retrieval, deterministic
recall_at_3	= 1.00	retrieval, deterministic
mean_faithfulness	≥ 0.9686	Claude judge, σ ≈ 0.005

Gated by .github/workflows/benchmark.yml. Faithfulness gating engages when the PR touches retrieval, reranker, expander, pipeline, prompts, or eval paths, or when the PR title contains [full-bench]. Methodology + raw numbers in docs/specs/release-quality-baseline/.

Per-hot-path latency

Locked dual-axis (wall-clock + CPU-time) thresholds on the four benchmarks. CPU-time is the gating axis (deterministic); wall-clock is advisory.

Numbers measured under the V2 multi-run methodology (5 invocations × 20 runs = 100 measurements per metric) on the locked-baseline runner (Linux ubuntu-latest, CPython 3.11.15). Inter-run and intra-run variance are tracked separately; thresholds are mean + 2σ × inter_run_stdev. Full 8-row dual-axis table + hardware fingerprint + per-metric noise profile: docs/specs/downstream-validation/perf-baseline.md.

Why two threshold styles in the locked table:

• keyword_retriever_retrieve has a wider CPU band because measured intra-run variance reflects cold-cache effects on the first few iterations — empirically derived, not tuned for tightness.
• llm_reranker_rerank is wall-clock-only because Anthropic network variance dominates the CPU axis; the gate is set generously.

Gated by .github/workflows/perf.yml per-PR (blocking on the CPU axis as of W3.1).

Why this is the differentiator

Most RAG libraries A/B-test internally and ship the result. attune-rag publishes the thresholds, gates merges against them, and re-measures whenever the corpus, judge prompt, or hardware changes. The receipts are checked in.

Bundled .help/ corpus

The repo ships a polished .help/ corpus that documents attune-rag's own surface — 143 templates across 13 features × 11 kinds (concept, task, reference, quickstart, faq, error, warning, tip, note, comparison, troubleshooting). Generated by attune-author with strict fact-check; queryable via AttuneHelpCorpus or as the bundled default for RagPipeline(). See .help/features.yaml for the feature map and .help/templates/ for the content.

The 13 features: pipeline, retrieval, corpus, prompts, provenance, providers, eval, benchmark, cli, editor, dashboard, expander, reranker.

What faithfulness measures

Faithfulness scores how well an answer is grounded in the retrieved passages — 1.0 means every claim in the answer is supported by a cited source; lower scores mean some claims have no support in the context. It catches hallucination in a way that precision_at_k and recall_at_k can't: those only measure whether the right documents were retrieved, not whether the generated answer actually used them.

attune-rag uses Claude as the judge via Anthropic's tool-use API to produce a structured score in [0.0, 1.0] for each (query, answer, retrieved_context) triple. The reported metric is the mean over the golden query set. Aggregate σ ≈ 0.005 over 40 queries even though per-query judge non-determinism can swing 40+ percentage points on individual queries — averaging absorbs the noise.

The same discipline powers attune-author's polish/fact-check pipeline — generated help content is scored against retrieved source passages before being marked authoritative. attune-rag's faithfulness primitives aren't just instrumentation; they're the contract the family's content-quality story is built on.

Run faithfulness manually

pip install 'attune-rag[claude]'
export ANTHROPIC_API_KEY=sk-ant-...

# Retrieval metrics only (free, deterministic):
attune-rag-benchmark --queries queries.yaml --json out.json

# Add faithfulness (~1 Claude API call per query, costs tokens):
attune-rag-benchmark --queries queries.yaml --with-faithfulness --json out.json

# Compare extended-thinking on vs off (2× judge cost):
attune-rag-benchmark --queries queries.yaml --with-faithfulness --compare-thinking --json out.json

The judge implementation lives at attune_rag.eval.faithfulness.FaithfulnessJudge. Note: attune_rag.eval.* is currently INTERNAL and may move — the attune-rag-benchmark --with-faithfulness CLI is the stable contract.

For the methodology behind the 0.9686 threshold, the v1/v2 ground-truth calibration runs, and the extended-thinking-vs-default decision record, see docs/rag/faithfulness-thinking-calibration.md.

Roadmap — embeddings (post-freeze 0.2.0+)

Keyword retrieval + optional Claude reranker currently meet the locked P@1 ≥ 0.95, R@3 = 1.00 thresholds against the attune-help golden set. The remaining hard queries (3 of 28, currently xpass-gated under [no-embeddings]) have zero token overlap against their target doc (e.g. "vulnerability scan" → tool-security-audit.md). Closing that gap needs vector search.

The plan is to ship attune-rag[embeddings] using fastembed for local, CPU-only embeddings — no new network dependency, no API key required at retrieval time. Keyword retrieval stays the default; embeddings layer in opt-in, same shape as QueryExpander and LLMReranker. With 0.2.0 cut, embeddings are a Phase 5 candidate — see docs/specs/ROADMAP-v1.md.

See CHANGELOG.md for the decision record and remaining-gap analysis.

Prompt caching (Claude only)

When using the Claude provider, run_and_generate automatically enables Anthropic prompt caching on the stable RAG context prefix (≥ 1 024 chars). This eliminates repeated token costs on the corpus portion of the prompt when the same context block is reused across calls.

No configuration needed — the provider handles the cache_control header automatically.

Public API

attune-rag's public surface is documented below and snapshot-tested in tests/unit/test_api_surface.py. Formal SemVer commitments are in effect as of 0.2.0 — see docs/POLICY.md for the deprecation policy. Symbols PUBLIC in 0.2.x stay PUBLIC through every 0.2.z; the snapshot test catches drift.

Top-level (from attune_rag import ...):

• Pipeline — RagPipeline, RagResult
• Corpus — CorpusProtocol, RetrievalEntry, DirectoryCorpus, AttuneHelpCorpus
• Retrieval — KeywordRetriever, RetrievalHit, RetrieverProtocol
• Provenance — CitationRecord, CitedSource, ClaimCitation, format_citations_markdown, format_claim_citations_markdown
• Prompting — build_augmented_prompt, PROMPT_VARIANTS
• Hybrid retrieval — QueryExpander, LLMReranker

PUBLIC submodules (importable by qualified path):

• attune_rag.corpus — exposes AliasInfo, DuplicateAliasError, load_aliases_from_file in addition to the top-level corpus names
• attune_rag.corpus.attune_help — AttuneHelpCorpus
• attune_rag.corpus.help_adapter — HelpCorpusAdapter Protocol
• attune_rag.providers — LLMProvider, get_provider, list_available
• attune_rag.measure_corpus — measure(...) function + MeasureResult dataclass for scoring a corpus against a query set. CLI via python -m attune_rag.measure_corpus ... or the attune-rag-measure console script. See docs/USER_CORPUS_GUIDE.md §6 for the worked example.
• attune_rag.editor — template-editor primitives (lint, schema, rename, autocomplete, references); see "Template editor primitives" above for the symbol list
• attune_rag.editor.{rename,schema,lint,autocomplete,references} — the individual editor submodules

Console scripts:

• attune-rag — CLI entry point (attune_rag.cli:main)
• attune-rag-measure — quality measurement (attune_rag.measure_corpus:main); CI-suitable via --watermark-r3 (non-zero exit on fail)

Anything not listed above is INTERNAL and may change in any release. The underscore-prefixed editor modules (attune_rag.editor._rename etc.) shipped in 0.1.x are deprecation shims as of 0.2.0; they re-export the new non-underscore names and emit DeprecationWarning. They are removed in 0.3.0.

Status

0.2.0 — first SemVer-binding cut. Phase 4 of the v1.0 roadmap landed cleanly: quality baselines (P@1 ≥ 0.95, R@3 = 1.00, mean faithfulness ≥ 0.9686) hold; per-hot-path perf thresholds re-locked under the V2 multi-run methodology (5 × 20 measurements); attune-gui downstream blocking gate stayed green throughout. From 0.2.0 forward, docs/POLICY.md §2 binds — symbols PUBLIC in 0.2.x stay PUBLIC through every 0.2.z.

We hit our Phase 4 goals ~3 weeks ahead of the nominal calendar and opted to ship early via the freeze-override mechanism rather than let the cadence clock run out — getting the user-facing additions (attune-rag-measure console script + attune_rag.measure_corpus module for benchmarking your own corpus quality; load_aliases_from_file() for file-based alias customization) into your hands sooner. Override rationale + per-PR receipts at docs/specs/api-v0.2.0-cut/.

Classifier stays at 3 - Alpha — the Production/Stable flip is a Phase 5 deliverable.

Part of the attune ecosystem (attune-ai, attune-help, attune-author, attune-gui).

License

Apache 2.0. See LICENSE.