piragi

The best RAG interface yet.

from piragi import Ragi

kb = Ragi(["./docs", "s3://bucket/data/**/*.pdf", "https://api.example.com/docs"])
answer = kb.ask("How do I deploy this?")

Built-in vector store, embeddings, citations, and auto-updates. Free & local by default. Ranked #1 Overall on the live AIR-Bench 24.05 Long-Doc Leaderboard (84.96% Recall@10, 75.72% nDCG@10, 80.10% MRR@10) and 32.61% Macro NDCG@10 (2.87x BM25; 71.98% Recall@100) across all 12 splits of BRIGHT (xlangai/BRIGHT, ICLR 2025) (31.43% fully open-weights — details).

Installation

pip install piragi

# Optional: Install Ollama for local LLM
curl -fsSL https://ollama.com/install.sh | sh
ollama pull llama3.2

# Optional extras
pip install piragi[s3]       # S3 support
pip install piragi[gcs]      # Google Cloud Storage
pip install piragi[azure]    # Azure Blob Storage
pip install piragi[crawler]  # Recursive web crawling
pip install piragi[graph]    # Knowledge graph
pip install piragi[postgres] # PostgreSQL/pgvector
pip install piragi[pinecone] # Pinecone
pip install piragi[supabase] # Supabase
pip install piragi[all]      # Everything

Features

• Zero Config - Works with free local models out of the box
• All Formats - PDF, Word, Excel, Markdown, Code, URLs, Images, Audio
• Remote Storage - Read from S3, GCS, Azure, HDFS, SFTP with glob patterns
• Web Crawling - Recursively crawl websites with /** syntax
• Auto-Updates - Background refresh, queries never blocked
• Smart Citations - Every answer includes sources
• Pluggable Stores - LanceDB, PostgreSQL, Pinecone, Supabase, or custom
• Advanced Retrieval - HyDE, hybrid search, cross-encoder reranking
• Semantic Chunking - Context-aware and hierarchical chunking
• Knowledge Graph - Entity/relationship extraction for better answers
• Async Support - Non-blocking API for web frameworks

Quick Start

from piragi import Ragi

# Local files
kb = Ragi("./docs")

# Multiple sources with globs
kb = Ragi(["./docs/*.pdf", "https://api.docs.com", "./code/**/*.py"])

# Remote filesystems
kb = Ragi("s3://bucket/docs/**/*.pdf")
kb = Ragi("gs://bucket/reports/*.md")

# Ask questions
answer = kb.ask("What is the API rate limit?")
print(answer.text)

# View citations
for cite in answer.citations:
    print(f"{cite.source}: {cite.score:.0%}")

Remote Filesystems

Read files from cloud storage using glob patterns:

# S3
kb = Ragi("s3://my-bucket/docs/**/*.pdf")

# Google Cloud Storage
kb = Ragi("gs://my-bucket/reports/*.md")

# Azure Blob Storage
kb = Ragi("az://my-container/files/*.txt")

# Mix local and remote
kb = Ragi([
    "./local-docs",
    "s3://bucket/remote-docs/**/*.pdf",
    "https://example.com/api-docs"
])

Requires optional extras: pip install piragi[s3], piragi[gcs], or piragi[azure]

Web Crawling

Recursively crawl websites using /** suffix:

# Crawl entire site
kb = Ragi("https://docs.example.com/**")

# Crawl specific section
kb = Ragi("https://docs.example.com/api/**")

# Mix with other sources
kb = Ragi([
    "./local-docs",
    "https://docs.example.com/**",
    "s3://bucket/data/*.pdf"
])

Crawls same-domain links up to depth 3, max 100 pages by default.

Requires: pip install piragi[crawler]

Vector Store Backends

from piragi import Ragi
from piragi.stores import PineconeStore, SupabaseStore

# LanceDB (default) - local or S3-backed
kb = Ragi("./docs")
kb = Ragi("./docs", store="s3://bucket/indices")

# PostgreSQL with pgvector
kb = Ragi("./docs", store="postgres://user:pass@localhost/db")

# Pinecone
kb = Ragi("./docs", store=PineconeStore(api_key="...", index_name="my-index"))

# Supabase
kb = Ragi("./docs", store=SupabaseStore(url="https://xxx.supabase.co", key="..."))

Advanced Retrieval

kb = Ragi("./docs", config={
    "retrieval": {
        "use_hyde": True,                  # Hypothetical document embeddings
        "use_hybrid_search": True,         # BM25 + vector search (RRF)
        "use_cross_encoder": True,         # Neural reranking
    }
})

# Reasoning-Intensive Retrieval (BRIGHT-grade multi-hop theorem/code search)
kb_reason = Ragi("./docs", config={
    "retrieval": {
        "use_think_then_retrieve": True,   # Pre-retrieval CoT decomposition
        "use_hybrid_search": True,
        "use_rrf": True,
        "use_cross_encoder": True,
        "reranker_type": "causal_llm",     # "cross_encoder" | "causal_llm" | "listwise"
        "cross_encoder_model": "Qwen/Qwen3-Reranker-4B",
        "rerank_instruction": "Retrieve passages providing foundational theorems or APIs.",
    }
})

Local Evidence Decisions (SemIf / OpenJev)

Optionally judge retrieved passages before the answer model sees them:

kb = Ragi("./docs", config={"decision": {"engine": "openjev"}})

Requires a separate SemIf installation. Piragi selects the device and downloads/caches a pinned model on the first decision. model_path is an optional offline override. Piragi retrieves a wider hybrid candidate pool, batches relevance judgments, then checks whether the retained passages support an answer. Insufficient evidence skips generation; inference errors fall back to ordinary retrieval and appear in answer.decision. Scores are not calibrated confidence. See setup, behavior, and evaluation.

Chunking Strategies

# Semantic - splits at topic boundaries
kb = Ragi("./docs", config={"chunk": {"strategy": "semantic"}})

# Hierarchical - parent-child for context + precision
kb = Ragi("./docs", config={"chunk": {"strategy": "hierarchical"}})

# Contextual - LLM-generated context per chunk
kb = Ragi("./docs", config={"chunk": {"strategy": "contextual"}})

Knowledge Graph

Extract entities and relationships for better multi-hop reasoning:

# Enable with single flag
kb = Ragi("./docs", graph=True)

# Automatic - extracts entities/relationships during ingestion
# Uses them to augment retrieval for relationship questions
answer = kb.ask("Who reports to Alice?")

# Direct graph access
kb.graph.entities()           # ["alice", "bob", "project x"]
kb.graph.neighbors("alice")   # ["bob", "engineering team"]
kb.graph.triples()            # [("alice", "manages", "bob"), ...]

Requires: pip install piragi[graph]

Configuration

from piragi import Ragi, RagiConfig

# Pydantic validation catches typos and invalid types
config = RagiConfig(llm={"model": "gpt-4", "api_key": "sk-..."})
kb = Ragi("./docs", config=config)

Configuration supports typed fields and automatic validation.

Async Support

Use AsyncRagi for non-blocking operations in async web frameworks:

from piragi import AsyncRagi

kb = AsyncRagi("./docs")

# Simple await
await kb.add("./more-docs")
answer = await kb.ask("What is X?")

# With progress tracking
async for progress in kb.add("./large-docs", progress=True):
    print(progress)
    # "Discovering files..."
    # "Found 10 documents"
    # "Chunking 1/10: doc1.md"
    # ...
    # "Generating embeddings for 150 chunks..."
    # "Embedded 32/150 chunks"
    # "Embedded 64/150 chunks"
    # ...
    # "Embeddings complete"
    # "Done"

# With FastAPI
@app.post("/ingest")
async def ingest(files: list[str]):
    await kb.add(files)
    return {"status": "done"}

All methods are async: add(), ask(), retrieve(), refresh(), count(), clear().

Retrieval Only

Use piragi as a retrieval layer without LLM:

chunks = kb.retrieve("How does auth work?", top_k=5)
for chunk in chunks:
    print(chunk.chunk, chunk.source, chunk.score)

# Use with your own LLM
context = "\n".join(c.chunk for c in chunks)
response = your_llm(f"Context:\n{context}\n\nQuestion: {query}")

LLM Client

from piragi import LLMClient

# Standalone unified client with connection pooling and backoff retries
client = LLMClient(model="gpt-4", api_key="sk-...")
response = client.generate("Hello!")

Pipelines & Advanced API

from piragi.pipelines import IngestionPipeline, RetrievalPipeline
from piragi.loader import DocumentLoader

# Composable pipelines for power users
ingest = IngestionPipeline(store=store, embedder=embedder)
# Memory-efficient streaming loader for large corpora
ingest.run(DocumentLoader.stream("./large-corpora/**/*.pdf")) 

# Parallel multi-query search via ThreadPoolExecutor
retrieve = RetrievalPipeline(store=store, llm=llm, max_parallel_queries=4)
retrieve.run("What is X?") 

Pipelines offer fine-grained control over ingestion and parallelized retrieval.

Observability

Built-in RED/USE metrics (p50/p95/p99), W3C distributed traces, structured JSON logs, and symptom alerts — zero setup:

stats = kb.stats()
print(stats["red"]["ask"])   # {'requests': 12, 'errors': 0, 'p50_ms': 41.2, 'p99_ms': 89.4, ...}
print(stats["alerts"])       # Active symptom-based alerts linked to docs/runbooks/

API

# Sync API
kb = Ragi(sources, persist_dir=".piragi", config=None, store=None, graph=False)
kb.add("./more-docs")
text, citations = kb.ask(query, top_k=5)
kb.retrieve(query, top_k=5)
kb.filter(**metadata).ask(query)  # Thread-safe filtering returns FilteredRagi
kb.refresh("./docs")
kb.count()
kb.stats()
kb.clear()

# Async API (same methods, just await them)
kb = AsyncRagi(sources, persist_dir=".piragi", config=None, store=None, graph=False)
await kb.add("./more-docs")
text, citations = await kb.ask(query, top_k=5)
await kb.stats()

Full docs: API.md

Playground

Interactive playground with zero setup — works in demo mode on GitHub Pages or live with a local backend:

pip install piragi
piragi playground

Features: multi-cell code editor, simulated outputs, LLM config, file browser.

License

MIT