consistent-rag 0.1.6

ConsistentRAG: Improving factual consistency in RAG through knowledge graph grounding and multi-agent refinement

ConsistentRAG

ConsistentRAG: Improving Factual Consistency in RAG via Graph-Based Retrieval and Ranking

A modular Python framework that improves factual consistency in Retrieval-Augmented Generation by grounding queries in a live knowledge graph. It uses a suite of graph algorithms to retrieve and rank structurally sound reasoning paths, which are then refined through a multi-agent loop where specialized AI "critics" collaboratively improve the answer across multiple iterations.

---

Try It in 5 Minutes (No Infrastructure Needed)

The fastest way to see ConsistentRAG in action uses KG-only mode — no vector database, no Docker, just Python + an LLM API key.

# 1. Clone and install
git clone https://github.com/pinkfloydsito/consistent-rag.git
cd consistent-rag
uv sync

# 2. Set your API key (pick one)
export OPENAI_API_KEY="sk-..."           # OpenAI, Azure, or any OpenAI-compatible endpoint
# OR
export DEEPSEEK_API_KEY="sk-..."         # DeepSeek (cheaper, no credit card needed)

# 3. Run the simplest example
uv run python examples/kg_rag_example.py

That's it. The example builds a knowledge graph from text and answers questions using graph reasoning.

---

Prerequisites

API Keys (Required)

ConsistentRAG needs an LLM for extraction, answering, and critique. Set at least one of these:

Priority	Variable	Example	Notes
1st	OPENAI_API_KEY	sk-...	OpenAI, Azure, or any OpenAI-compatible provider
2nd	DEEPSEEK_API_KEY	sk-...	Fallback if OpenAI key is absent
—	OPENAI_API_BASE	https://api.openai.com/v1	Override the API endpoint (e.g., for Azure or local LLMs)

Copy .env.example to .env and fill in your keys:

cp .env.example .env
# Edit .env with your keys

Infrastructure (Only for specific modes)

What	When needed	How to start
Qdrant	Online modes (backend="qdrant")	make docker-up
PostgreSQL	Experiment tracking / integration tests	make docker-up (same command)

The make docker-up command starts both Qdrant (port 6333) and PostgreSQL (port 5432). You don't need them for KG-only or FAISS-offline modes.

---

Installation

# With uv (recommended)
uv sync --extra dev

# With pip
pip install -e ".[dev]"

---

Quick Start: Three Paths

Path A: "Just want to try it" — KG-Only, No Servers

No vector DB. No Docker. Just a graph built from text.

from consistent_rag import ConsistentRAG

rag = ConsistentRAG(pipeline="kg_only", strategy="ppr")

context = """
TechStart was founded in 2019 by Alice Johnson in San Francisco.
The company focuses on AI-powered analytics for retail businesses.
"""

result = rag.query("Who founded TechStart?", context=context)
print(result.answer)
# Output: "TechStart was founded by Alice Johnson."

What's happening: The pipeline extracts triples from text → builds a NetworkX graph → uses PPR to rank reasoning paths → generates an answer.

See examples/kg_rag_example.py for a full runnable script.

Path B: "I have documents to index" — FAISS + KG, Fully Local

Good for: offline use, no external services, saving/loading indices.

from consistent_rag import ConsistentRAG

rag = ConsistentRAG(
    pipeline="hybrid",
    backend="faiss",
    strategy="ppr",
    embedding_model="all-MiniLM-L6-v2",  # local embeddings, no API key
)

# Index your documents
docs = [
    "Pydantic AI is a Python agent framework...",
    "DeepSeek develops large language models...",
]
rag.index_documents(docs, recreate=True, build_kg=True)

# Query
result = rag.query("What is Pydantic AI?")
print(result.answer)

# Save for later
rag.save_index("./my_index")

See examples/agentic_rag_example.py for a complete example.

Path C: "Production setup" — Qdrant + Full Pipeline

Good for: large document collections, online dynamic KG, adaptive strategies.

# 1. Start infrastructure
make docker-up

from consistent_rag import ConsistentRAG

rag = ConsistentRAG(
    pipeline="hybrid",
    backend="qdrant",
    collection_name="my_docs",
    strategy="adaptive",  # auto-selects best strategy per query
)

# Index documents (KG built incrementally)
rag.index_documents(docs, build_kg=True)

# Query with full agentic loop
result = rag.query("What is the relationship between X and Y?")

print(f"Answer: {result.answer}")
print(f"Iterations: {result.iterations_used}")
print(f"Critic score: {result.final_score:.2f}")
print(f"Seeds: {result.seeds_used}")

See examples/basic_rag_example.py for vector-only mode and examples/kg_ppr_rag_example.py for KG+PPR with Qdrant.

---

Streamlit Visualization

Run the interactive demo to watch the pipeline execute step-by-step:

uv run streamlit run streamlit_app.py

Features:

• Live query input with strategy selection
• Per-iteration subgraph visualization
• Critic scores, weight deltas, and convergence tracking
• Benchmark dataset selection (FaithEval, SQuAD, etc.)
• Reasoning path evolution charts

---

Configuration

All parameters are passed to ConsistentRAG() or PipelineConfig:

from consistent_rag import ConsistentRAG

rag = ConsistentRAG(
    # Pipeline mode
    pipeline="hybrid",          # "vector_only", "kg_only", or "hybrid"
    backend="faiss",            # "faiss" or "qdrant" (vector pipelines only)
    strategy="ppr",             # "ppr", "nhops", "random_walk", "hybrid", "adaptive"

    # Model settings
    llm_model="gpt-4o-mini",    # or "deepseek-chat", etc.
    embedding_model="all-MiniLM-L6-v2",

    # Agentic loop
    max_iterations=3,           # Max answer-critic iterations
    improvement_threshold=0.8,  # Stop if critic score >= this

    # Graph algorithm
    ppr_alpha=0.5,              # PPR damping factor
    max_paths=20,               # Max reasoning paths in context
    path_max_hops=3,            # Max hops per path
)

See consistent_rag/pipeline/config.py for the full parameter list.

---

Pipeline Configurations

Configuration	Type	Vector	KG	Graph Strategy
baseline_no_tool	Single-pass	No	No	—
baseline_vector	Single-pass	Yes	No	—
baseline_kg	Single-pass	No	Yes	PPR
baseline_hybrid	Single-pass	Yes	Yes	PPR
agentic_vector_only	Multi-agent	Yes	No	—
agentic_kg_nhops	Multi-agent	Yes	Yes	N-Hops BFS
agentic_kg_ppr	Multi-agent	Yes	Yes	PPR
agentic_kg_random_walk	Multi-agent	Yes	Yes	Random Walk
agentic_kg_hybrid	Multi-agent	Yes	Yes	Hybrid Semantic
agentic_kg_adaptive	Multi-agent	Yes	Yes	Adaptive (Full)

Run any configuration via the evaluation CLI:

uv run python -m consistent_rag.evaluate_all \
    --approaches agentic_kg_ppr \
    --benchmark faitheval \
    --limit 10 --verbose

---

Architecture

ConsistentRAG operates as a four-phase pipeline:

+-------------------------------------------------------------+
|                     Pipeline Layer                          |
| (AdaptiveHybrid / OnlineDynamic / OfflineKG / Baseline)     |
+-------------------------------------------------------------+
|                   Orchestration Layer                        |
|    (AdaptiveRouter Orchestrator / Multi-Agent Loop)          |
+-------------------------------------------------------------+
|                      Agent Layer                            |
|   Seed Agent | Router Agent | Answer Agent | Critic         |
+-------------------------------------------------------------+
|                  Core Services Layer                        |
|   Retriever | Graph Store | Multi-Algorithm Engine | LLM    |
+-------------------------------------------------------------+
|                  Infrastructure Layer                       |
|   Qdrant | NetworkX | FAISS | DeepSeek API (Default)        |
|                       | OpenAI API (Configurable)           |
+-------------------------------------------------------------+

Four Phases

1. Dual-Mode Indexing & KG Construction — Offline (FAISS + pre-built graph) or Online (Qdrant + on-the-fly graph)
2. Multi-Algorithmic Graph Retrieval Engine — PPR, Directed Random Walks, Hybrid Semantic Traversal, N-Hops BFS
3. Adaptive Strategy Routing — Router Agent dynamically selects the optimal graph algorithm per query
4. Agentic Iterative Refinement — Answer + Critic loop with context pruning, structural expansion, and instructional augmentation

---

Running Evaluations

The evaluation system runs any combination of the 10 pipeline configurations against the 5 benchmarks. Results are persisted to CSV (and optionally PostgreSQL) with full resume support.

# Quick smoke test: 1 config, 1 benchmark, 2 samples
uv run python -m consistent_rag.evaluate_all \
    --approaches baseline_no_tool \
    --benchmark faitheval \
    --limit 2 --verbose

# Run a specific config against all benchmarks
uv run python -m consistent_rag.evaluate_all \
    --approaches agentic_kg_ppr \
    --benchmark all \
    --limit 50 \
    --csv results/agentic_kg_ppr.csv

# Run all configs against all benchmarks (full thesis experiment matrix)
uv run python -m consistent_rag.evaluate_all \
    --approaches all \
    --benchmark all \
    --csv results/full_matrix.csv \
    --metrics all

# Resume an interrupted run
uv run python -m consistent_rag.evaluate_all \
    --approaches all \
    --benchmark all \
    --csv results/full_matrix.csv \
    --resume

# With PostgreSQL persistence
uv run python -m consistent_rag.evaluate_all \
    --approaches all \
    --benchmark faitheval \
    --csv results/faitheval.csv \
    --postgres postgresql://consistent_rag:consistent_rag@localhost:5432/consistent_rag

CLI Options

Flag	Description
--approaches	Comma-separated config names or all
--benchmark	Comma-separated benchmark names or all
--limit	Max samples per benchmark (default: 5)
--metrics	basic, llm, deepeval, or all
--csv	Output CSV path (auto-generated if omitted)
--resume	Skip completed samples in existing CSV
--no-cache	Re-extract triplets instead of using cache
--postgres	PostgreSQL URL for experiment tracking
--verbose	Print per-sample progress

Build Offline Index

uv run python scripts/build_offline_index.py

---

Benchmarks

Benchmark	Focus	Samples
FaithEval	Factual consistency (unanswerable, inconsistent, counterfactual)	~15K
MuSiQuE	Multi-hop reasoning across documents	~25K
TimeQA	Temporal reasoning and evolving facts	~20K
SQuAD	Single-hop extractive QA (control baseline)	~100K
FinanceBench	Domain-specific financial document QA (SEC filings)	150

---

MCP Tool

ConsistentRAG can be used as an MCP tool by any compatible agent:

# Start the MCP server
python -m consistent_rag.mcp_server

The server exposes a query tool with parameters:

• question (required): The question to answer
• context (optional): Pre-provided context (skips retrieval)
• mode: online_dynamic or offline_static
• strategy: adaptive, ppr, random_walk, hybrid, nhops
• top_k: Number of documents to retrieve

---

Troubleshooting

"No API key found"

ValueError: No OpenAI or DeepSeek API key found

Fix: Set OPENAI_API_KEY or DEEPSEEK_API_KEY in your environment or .env file.

"Connection refused" to localhost:6333

qdrant_client.http.exceptions.ResponseHandlingException

Fix: Qdrant is not running. Start it with make docker-up. If you're using FAISS or KG-only mode, you don't need Qdrant.

"Module not found" after installation

Fix: Make sure you're in the correct virtual environment:

uv sync --extra dev
source .venv/bin/activate  # or .venv\Scripts\activate on Windows

DeepSeek format errors

DeepSeek occasionally rejects message formats. The pipeline automatically retries once. If it persists, switch to OpenAI or another provider.

Out of memory during indexing

For large document collections, reduce max_triples_per_doc (default: 40) or use Qdrant instead of FAISS:

rag = ConsistentRAG(
    pipeline="hybrid",
    backend="qdrant",  # better for large collections
    max_triples_per_doc=20,
)

---

Development

make lint          # Format check + linting
make format        # Auto-fix formatting
make test-unit     # Run unit tests (no external services)
make test          # Run all tests
make ci            # Full CI pipeline
make docker-up     # Start Qdrant + PostgreSQL
make docker-down   # Stop infrastructure

---

Project Structure

consistent_rag/
├── api.py                       # Public API (ConsistentRAG, QueryResult)
├── pipeline/
│   ├── config.py                # PipelineConfig (unified configuration)
│   ├── pipeline.py              # Main pipeline orchestrator
│   └── pipeline_factory.py      # Pipeline configuration factory
├── llm.py                       # Universal LLM client (OpenAI-compatible)
├── embeddings.py                # Embedding service (local + API)
├── retrievers/                  # Vector retrievers (Qdrant, FAISS)
├── knowledge_graph/
│   ├── networkx_graph_store.py  # NetworkX graph backend
│   ├── ppr_engine.py            # Personalized PageRank
│   ├── algorithms.py            # Random Walk, Hybrid, N-Hops
│   ├── extractor.py             # LLM-based triple extraction
│   └── lsh_index.py             # Locality Sensitive Hashing
├── agents/
│   ├── answer/                  # Answer generation agent
│   ├── critic/                  # Critic agent with structured feedback
│   └── retrieval/               # Retrieval orchestrator
├── benchmarks/                  # Dataset loaders
├── evaluation/                  # Metrics and experiment runner
└── examples/                    # Runnable example scripts