agentic-rag-core 2.1.2

Vectorless, Reasoning-based RAG — multi-document agentic retrieval with zero hallucination

AgenticRAG

Vectorless, Reasoning-based RAG for Python

No Vector DB  ·  No Chunking  ·  No Embeddings
Just pure LLM reasoning over your documents.

Installation  ·  Quick Start  ·  Web UI  ·  API Reference  ·  Models

---

AgenticRAG is a Python library that lets you ask questions about your PDF documents using AI — without any vector databases, embeddings, or chunking.

Instead of the traditional RAG approach, AgenticRAG:

1. Builds a smart tree index from your document (like a Table of Contents, but smarter)
2. Uses AI agents to reason over the tree and find the right sections
3. Verifies every answer against the source text — zero hallucinations

It works with Google Gemini, Groq Cloud (free API key) or local LLMs via Ollama (100% free, runs on your machine).

---

Installation

pip install agentic-rag-core

That's it. You're ready to go.

Optional extras:

pip install agentic-rag-core[web]    # Web UI (includes FastAPI server)
pip install agentic-rag-core[gcs]    # Google Cloud Storage backend
pip install agentic-rag-core[neo4j]  # Neo4j graph backend
pip install agentic-rag-core[all]    # Everything

Note: If you clone the repo and want to run server.py, install with: pip install -e ".[web]"

---

Quick Start (30 seconds)

Step 1: Get a free API key

Go to console.groq.com -- Create an API Key -- Copy it.

Step 2: Set your API key

Create a .env file in your project folder:

GROQ_API_KEY=gsk_your_key_here

[!NOTE] Using the Web UI? If you are running the browser-based Web UI, creating a .env file is entirely optional! You can set your API keys directly inside the interface's settings panel.

Step 3: Ask questions about any PDF

from agenticrag import Forest

# Create a knowledge base and add your PDF
forest = Forest(verbose=True)
forest.add("report.pdf")

# Ask a question
result = forest.ask("What was the net income?")
print(result.text)

That's the entire setup. Three lines of real code. No vector DB to configure, no embeddings to generate, no chunks to tune.

Or: Run the Web UI (Zero Coding!)

If you prefer a beautiful, browser-based chat interface over writing Python scripts, you can spin up the Web UI directly:

# 1. Install with Web UI dependencies
pip install agentic-rag-core[web]

# 2. Start the web server
python -m agenticrag serve

Open http://localhost:8000 in your browser to start uploading PDFs and chatting!

---

Use AgenticRAG in Your Own Project

AgenticRAG is designed to be a drop-in RAG engine for any Python project. You don't need to understand how RAG works, build any retrieval pipelines, or set up any infrastructure. Just install, import, and ask questions.

How It Works (The Simple Version)

Your app  -->  agenticrag  -->  Answer with sources
               (handles everything:
                PDF parsing, indexing, multi-agent search,
                hallucination checking, citations)

You write zero retrieval code. AgenticRAG handles all of it internally.

Example 1: Add Document Q&A to a Flask App

# app.py
from flask import Flask, request, jsonify
from agenticrag import Forest

app = Flask(__name__)

# Create the knowledge base ONCE when the app starts
forest = Forest(verbose=True)
forest.add("company_handbook.pdf")
forest.add("product_docs.pdf")

@app.route("/ask", methods=["POST"])
def ask():
    question = request.json["question"]
    result = forest.ask(question)
    return jsonify({
        "answer": result.text,
        "confidence": result.confidence,
        "sources": result.sources,
    })

if __name__ == "__main__":
    app.run(port=5000)

# Install
pip install agentic-rag-core flask

# Run
python app.py

# Test
curl -X POST http://localhost:5000/ask \
  -H "Content-Type: application/json" \
  -d '{"question": "What is our refund policy?"}'

Example 2: Build a FastAPI Document API

# api.py
from fastapi import FastAPI
from pydantic import BaseModel
from agenticrag import Forest

app = FastAPI(title="My Document API")
forest = Forest(data_dir="./knowledge_base", verbose=True)

# Add all your documents at startup
forest.add_directory("./docs/", pattern="*.pdf")

class Question(BaseModel):
    text: str

@app.post("/query")
async def query(q: Question):
    result = forest.ask(q.text)
    return {
        "answer": result.text,
        "confidence": result.confidence,
        "pages_used": result.sources,
        "time_seconds": result.elapsed_seconds,
    }

pip install agentic-rag-core fastapi uvicorn
uvicorn api:app --reload

Example 3: Streamlit Chat App (10 lines)

# chat.py
import streamlit as st
from agenticrag import Forest

st.title("Chat with your Documents")

# Initialize once
if "forest" not in st.session_state:
    st.session_state.forest = Forest(verbose=True)
    st.session_state.forest.add("report.pdf")

question = st.text_input("Ask a question:")
if question:
    result = st.session_state.forest.ask(question)
    st.write(result.text)
    st.caption(f"Confidence: {result.confidence:.0%} | {result.elapsed_seconds:.1f}s")

pip install agentic-rag-core streamlit
streamlit run chat.py

Example 4: Simple Python Script

The simplest possible usage — no web framework, no server, just a script:

# ask.py
from agenticrag import Forest

# Point to your documents
forest = Forest()
forest.add("quarterly_report.pdf")
forest.add("annual_report.pdf")

# Ask questions
questions = [
    "What was the total revenue?",
    "What are the main risk factors?",
    "How did expenses change year over year?",
]

for q in questions:
    result = forest.ask(q)
    print(f"\nQ: {q}")
    print(f"A: {result.text}")
    print(f"   Confidence: {result.confidence:.0%}")
    print(f"   Sources: {[s['doc_title'] + ' p.' + s['pages'] for s in result.sources]}")

pip install agentic-rag-core
python ask.py

Example 5: Add to an Existing Django Project

# views.py
from django.http import JsonResponse
from django.views.decorators.csrf import csrf_exempt
from agenticrag import Forest
import json

# Initialize once — reused across all requests
_forest = Forest(data_dir="./django_knowledge_base")
_forest.add_directory("./media/documents/", pattern="*.pdf")

@csrf_exempt
def document_qa(request):
    if request.method == "POST":
        body = json.loads(request.body)
        result = _forest.ask(body["question"])
        return JsonResponse({
            "answer": result.text,
            "confidence": result.confidence,
        })

Example 6: Use with Local LLMs (100% Free, No API Key)

from agenticrag import Forest, LocalModel

# No API key needed — runs on your machine
forest = Forest(
    model=LocalModel.QWEN3_4B,
    base_url="http://localhost:11434/v1",
)
forest.add("confidential_report.pdf")  # data never leaves your machine
result = forest.ask("What are the projected earnings?")
print(result.text)

Example 7: Build a CLI Tool

# doctool.py
import sys
from agenticrag import Forest

forest = Forest(data_dir="./my_index")

if len(sys.argv) > 2 and sys.argv[1] == "add":
    result = forest.add(sys.argv[2])
    print(f"Indexed: {result.title} ({result.page_count} pages)")

elif len(sys.argv) > 2 and sys.argv[1] == "ask":
    question = " ".join(sys.argv[2:])
    result = forest.ask(question)
    print(f"\n{result.text}\n")
    print(f"Confidence: {result.confidence:.0%} | Sources: {len(result.sources)}")

else:
    print("Usage:")
    print("  python doctool.py add report.pdf")
    print('  python doctool.py ask "What was the revenue?"')

python doctool.py add report.pdf
python doctool.py ask "What was the total revenue in Q4?"

The Key Idea

No matter what you're building — a web app, an API, a chatbot, a CLI tool, or a data pipeline — the pattern is always the same:

from agenticrag import Forest

# 1. Create a Forest (one time)
forest = Forest()

# 2. Add your documents (one time — they're cached)
forest.add("your_file.pdf")

# 3. Ask questions (as many times as you want)
result = forest.ask("your question here")
print(result.text)       # the answer
print(result.confidence) # how sure it is
print(result.sources)    # where it found the info

That's it. Three steps. No RAG architecture knowledge needed. AgenticRAG handles all the indexing, multi-agent search, synthesis, and hallucination checking internally.

---

Using AgenticRAG as a Library

Single Document — PageIndex

If you only have one document, use PageIndex for the simplest experience:

from agenticrag import PageIndex

# Load and index a PDF (one-time, takes ~30 seconds)
pi = PageIndex()
pi.load("annual_report.pdf")

# Ask questions
result = pi.ask("What was the net income in 2023?")
print(result.text)

# Save the index so you don't have to rebuild it next time
pi.save("report_index.json")

# Later, load it instantly
pi = PageIndex()
pi.load_json("report_index.json")
result = pi.ask("What about revenue?")

Multiple Documents — Forest

For multiple documents, use Forest. It searches across ALL your documents at once:

from agenticrag import Forest

# Create a knowledge base
forest = Forest(verbose=True)

# Add documents one by one
forest.add("report_2023.pdf")
forest.add("report_2024.pdf")

# Or add an entire folder of PDFs
forest.add_directory("./contracts/", pattern="*.pdf")

# Ask questions across ALL documents
result = forest.ask("Compare revenue growth between 2023 and 2024")

# Rich result object
print(result.text)               # The answer
print(result.confidence)         # 0.0-1.0 confidence score
print(result.sources)            # Which documents & pages were used
print(result.documents_searched) # Which doc IDs were searched
print(result.elapsed_seconds)    # How long it took

Using Local LLMs (Free, No API Key Needed)

You can run AgenticRAG 100% locally with Ollama:

# 1. Install Ollama from https://ollama.com/download
# 2. Pull a model (Qwen3 4B recommended — only 2.5 GB download)
ollama pull qwen3:4b

from agenticrag import Forest, LocalModel

forest = Forest(
    model=LocalModel.QWEN3_4B,
    base_url="http://localhost:11434/v1",
    verbose=True,
)
forest.add("report.pdf")
result = forest.ask("What are the key risks?")
print(result.text)

Batch Ingestion (100+ Documents)

For large collections, use the batch pipeline — it's 2-4x faster:

from agenticrag import Forest, LocalModel

forest = Forest(
    model=LocalModel.QWEN3_4B,
    base_url="http://localhost:11434/v1",
    data_dir="./my_knowledge_base",
    verbose=True,
)

# Ingest all PDFs with progress logging
result = forest.add_directory_batch(
    "./papers/",
    pattern="*.pdf",
    resume=True,              # skip already-indexed docs
    skip_description=True,    # halves LLM calls (faster)
    max_llm_concurrent=2,     # concurrent LLM requests
)

print(result)
# BatchResult(total=14000, succeeded=13950, failed=50, skipped=0)

# Now query across all documents
answer = forest.ask("What are the latest findings on X?")

Multi-Turn Conversations

AgenticRAG remembers your conversation automatically:

forest = Forest(verbose=True)
forest.add("report.pdf")

# First question
result = forest.ask("What was the revenue?")
print(result.text)  # "Revenue was $5.2 billion..."

# Follow-up — it remembers the context
result = forest.ask("How does that compare to last year?")
print(result.text)  # "Compared to last year's $4.8 billion..."

# Reset when you want to start fresh
forest.clear_history()

---

Web UI

AgenticRAG includes a beautiful web interface for chatting with your documents:

# 1. Install the library with web UI dependencies
pip install agentic-rag-core[web]

# 2. Start the web UI server
python -m agenticrag serve

# 3. Start on a custom port
python -m agenticrag serve --port 9000

This opens a web app at http://localhost:8000 where you can:

• Create notebooks to organize your documents
• Upload PDFs via drag-and-drop
• Chat with your documents (with source citations)
• Switch between Groq Cloud, Gemini, and local LLM providers
• Share over LAN — anyone on your network can access it

[!TIP] No .env file required! You can spin up the Web UI server immediately without creating a .env file or configuring system environment variables. Once the Web UI is running, simply click on the Settings panel in the top bar to choose your provider, enter your API key, or specify your local LLM's connection endpoint (e.g., Ollama). All settings are saved locally to your workspace.

Split Architecture (GPU on one machine, UI on another)

# Machine A (GPU): Run Ollama
set OLLAMA_HOST=0.0.0.0
ollama serve

# Machine B (laptop): Start UI and point AgenticRAG to Machine A
python -m agenticrag serve
# Then in Settings > Local LLM > Base URL: http://MACHINE_A_IP:11434/v1

---

API Reference

High-Level Classes

Class	What it does	When to use
Forest	Multi-document knowledge base	Most common — use this for everything
PageIndex	Single-document index	When you only have one document
ForestResult	Result from Forest.ask()	Access .text, .sources, .confidence

Forest Methods

Method	Description
Forest(model=..., verbose=True)	Create a new knowledge base
.add("file.pdf")	Add a single document
.add_directory("./docs/")	Add all PDFs from a folder
.add_directory_batch("./docs/")	Fast batch add (100+ docs)
.ask("question")	Ask a question across all docs
.documents()	List all indexed documents
.remove(doc_id)	Remove a document
.size	Number of documents
.clear_history()	Reset conversation memory
.info()	Forest status summary

PageIndex Methods

Method	Description
PageIndex(model=..., verbose=True)	Create a new single-doc index
.load("file.pdf")	Index a document
.save("index.json")	Save the index to disk
.load_json("index.json")	Load a saved index
.ask("question")	Ask a question

ForestResult Fields

Field	Type	Description
.text	str	The final verified answer
.confidence	float	0.0 to 1.0 confidence score
.sources	list	Which documents/pages were used
.documents_searched	list	Which doc IDs were searched
.reasoning_trace	list	Step-by-step agent pipeline trace
.was_rewritten	bool	Whether the Critic modified the answer
.hallucinations	list	Any hallucinations that were caught
.elapsed_seconds	float	Total time taken

Configuration

from agenticrag import ForestConfig, GroqModel

config = ForestConfig(
    model              = GroqModel.GPT_OSS_20B,  # Which AI model to use
    data_dir           = "./my_data",             # Where to store indices
    max_docs_per_query = 5,                       # Max docs to search per question
    max_hunt_workers   = 5,                       # Parallel search threads
    enable_critic      = True,                    # Zero-hallucination checking
    verbose            = True,                    # Print progress
    # ── Hybrid Sub-Tree Pre-Filtering ──────────────────────────────────
    # Activates for trees larger than pre_filter_threshold nodes.
    # One small KeywordAgent LLM call (256 tokens) replaces up to 10
    # large SELECT_NODES calls (25,000 tokens each) on big documents.
    enable_pre_filtering = True,   # Toggle the entire pre-filter pipeline
    pre_filter_threshold = 50,     # Min node count to activate (default: 50)
    max_filter_candidates = 20,    # Top-N matched seed nodes (default: 20)
)

Or pass these directly to Forest():

forest = Forest(
    model=GroqModel.GPT_OSS_20B,
    data_dir="./my_data",
    verbose=True,
    enable_pre_filtering=True,   # default — disable if you want full tree always
    pre_filter_threshold=50,     # lower = activates on smaller documents
    max_filter_candidates=20,    # higher = more candidate nodes, larger prompt
)

---

Supported Models

Cloud Models (Groq — Free API)

from agenticrag import Forest, GroqModel

forest = Forest(model=GroqModel.GPT_OSS_20B)      # Fast, recommended default
forest = Forest(model=GroqModel.GPT_OSS_120B)      # Largest, best reasoning
forest = Forest(model=GroqModel.LLAMA4_SCOUT)       # Llama 4 Scout
forest = Forest(model=GroqModel.LLAMA3_3_70B)       # Llama 3.3 70B
forest = Forest(model=GroqModel.QWEN3_32B)          # Qwen 3 32B
forest = Forest(model=GroqModel.DEEPSEEK_R1_DISTILL_LLAMA_70B)  # DeepSeek R1

Local Models (Ollama — 100% Free)

# Install from https://ollama.com/download, then:
ollama pull qwen3:4b     # 2.5 GB — recommended
ollama pull qwen3:8b     # 5.2 GB — better quality
ollama pull qwen3:14b    # 9.3 GB — even better

from agenticrag import Forest, LocalModel

forest = Forest(
    model=LocalModel.QWEN3_4B,
    base_url="http://localhost:11434/v1",
)

Model	Download Size	VRAM Needed	Best For
LocalModel.QWEN3_4B	2.5 GB	5 GB or less	Low-VRAM GPUs, fastest
LocalModel.QWEN3_8B	5.2 GB	8 GB or less	Best quality/size ratio
LocalModel.QWEN3_14B	9.3 GB	12 GB or less	Higher quality
LocalModel.QWEN3_30B	19 GB	24 GB or less	Strong reasoning
LocalModel.LLAMA3_2_3B	2.0 GB	4 GB or less	Ultra-lightweight
LocalModel.MISTRAL	4.1 GB	6 GB or less	General purpose
LocalModel.GEMMA3_12B	8.1 GB	12 GB or less	Alternative mid-range

---

How It Works

AgenticRAG uses a multi-agent pipeline — like a team of AI researchers working together:

Your Question
     |
     v
+---------+   Looks at the document graph to find
| Planner |-->  which documents might have the answer
+---------+
     |
     v
+---------+   Searches those documents IN PARALLEL
| Hunters |-->  using tree-based reasoning (not keywords!)
+---------+
     |
     v
+--------------+   Combines evidence from multiple docs
| Synthesizer  |-->  into a single, coherent answer
+--------------+
     |
     v
+--------+   Checks every claim against the source text
| Critic |-->  removes anything not backed by evidence
+--------+
     |
     v
  Verified Answer

This is why AgenticRAG can answer complex questions across many documents — it doesn't just find similar text, it actually reasons about what's relevant.

---

Comparison: AgenticRAG vs. Vector RAG

AgenticRAG is not meant to replace Vector RAG for all use cases. It is designed specifically for deep research on highly structured documents (like SEC filings, legal contracts, or technical manuals) where exact facts and numbers matter.

When to use Vector RAG

• Goal: Finding general semantic themes across millions of documents (e.g., "What is the company's general tone regarding AI?").
• Why: Vector RAG is fast, cheap, and scales to millions of documents instantly. It uses cosine similarity to find semantically related text.
• Weakness: It blindly chunks documents, often splitting tables or separating numbers from their context. It struggles with exact keyword constraints and is prone to LLM hallucinations when synthesizing answers from disjointed chunks.

When to use AgenticRAG

• Goal: Targeted, verifiable reasoning on a curated set of complex documents (e.g., "What was the exact amortisation of intangible assets in 2023?").
• Why:
  1. Context Preservation: It preserves the document's structure as a tree. Financial tables remain intact under their logical headings.
  2. Exact Retrieval: The KeywordAgent combined with regex pre-filtering ensures the system hunts for exact strings (like "Operating Margin") rather than mathematically similar concepts.
  3. Zero-Hallucination: The CriticAgent strictly cross-references the drafted answer against the raw retrieved text, removing any hallucinated numbers.
• Weakness: It is token-heavy and slow (multiple LLM calls per query). It is not built for high-throughput, low-latency web search over millions of documents.

---

Hybrid Sub-Tree Pre-Filtering

For large documents (SEC 10-K filings, legal contracts, technical manuals), the tree index can contain hundreds of nodes. Passing the entire tree into every SELECT_NODES call is slow, expensive, and causes the LLM to miss relevant nodes (needle-in-a-haystack problem).

AgenticRAG automatically activates Hybrid Sub-Tree Filtering when a tree exceeds pre_filter_threshold nodes: User Question │ ▼ KeywordAgent.expand() — 1 LLM call, 256 tokens max • Generates keyphrases, keywords, synonyms • Receives document_description so it predicts vocabulary specific to THIS document │ ▼ _local_node_search() — 0 LLM calls, pure Python regex • Applies Stem Title Bonus (+10 pts) for depluralised keyword matches • Scores exact hits across title, summary, and deep text-preview (2000 chars) • Returns node IDs ranked by hit count + bonuses │ ▼ _build_candidate_subtree() — 0 LLM calls • Keeps top-N matched nodes + all their ancestors • Preserves hierarchical context for the LLM │ ▼ SELECT_NODES loop — on ~15-node sub-tree, not 1,000-node full tree • ~500 tokens per call instead of ~25,000 tokens • 98% token reduction on large documents

Token savings example (3M 2018 10-K, ~160 pages)

Metric	Without Pre-Filter	With Pre-Filter
Nodes passed to LLM	~800	~15
Tokens per SELECT_NODES call	~20,000	~400
Total tokens (3 iterations)	~60,000	~1,200
Rate-limit risk	High	Negligible

Developer logging

Set verbose=True on your Forest to see the full pre-filter trace in the console. All steps are also written to pageindex_data/logs/trail.log:

==================== PRE-FILTER INITIATED ==================== INFO: Tree size (800 nodes) exceeds threshold (50). Running keyword pre-filtering to build a compact candidate sub-tree. ==================== KEYWORD AGENT (INPUT) ==================== DATA: { "question": "what distributions do we have?", "doc_context_preview": "This document is 3M Company's 2018 Annual Report..." } ==================== KEYWORD AGENT (SUCCESS) ==================== DATA: { "expanded_keywords": ["distributions", "dividends paid", "financing activities", "treasury stock", "stockholder payouts", ...] } ==================== PRE-FILTER COMPLETED ==================== INFO: Filtered 800 nodes → 18 candidate seed nodes + their ancestors. DATA: { "candidate_seed_ids": ["0042", "0043", "0089", ...], "matched_keywords": ["dividends paid", "financing activities", ...] }

Configuration reference

Parameter	Default	Description
enable_pre_filtering	True	Toggle the entire pipeline on/off
pre_filter_threshold	50	Min node count before activation
max_filter_candidates	20	Max seed nodes passed to sub-tree builder

When to tune these:

• Slow responses on small documents → raise pre_filter_threshold to 100+
• Missing relevant nodes on huge documents → raise max_filter_candidates to 30-40
• Pre-filter gives zero matches repeatedly → check trail.log for the expanded keywords

---

Project Structure

agenticrag/
|-- __init__.py              # Public API (Forest, PageIndex, etc.)
|-- __main__.py              # python -m agenticrag entry point
|-- config.py                # GroqModel, LocalModel, configuration
|-- groq_client.py           # LLM wrapper (Groq + OpenAI-compatible)
|-- pdf_parser.py            # PDF/Markdown text extraction
|-- pdf_to_markdown.py       # PDF to Markdown converter (no LLM)
|-- prompts.py               # All LLM prompts
|-- tree_builder.py          # Build hierarchical tree index
|-- tree_search.py           # Single-document tree search
|-- pageindex.py             # PageIndex (single-doc wrapper)
|-- forest.py                # Forest (multi-doc entry point)
|-- agents/
|   |-- planner.py           # Document selection from graph
|   |-- hunter.py            # Parallel document searching
|   |-- synthesizer.py       # Multi-doc answer synthesis
|   |-- evaluator.py         # Retrieval sufficiency checking
|   |-- critic.py            # Zero-hallucination enforcer
|   +-- orchestrator.py      # Agentic loop state machine
|-- storage/
|   |-- base.py              # Abstract TreeStore interface
|   |-- local.py             # Local filesystem (default)
|   +-- gcs.py               # Google Cloud Storage
|-- graph/
|   |-- base.py              # Abstract DocumentGraph interface
|   |-- sqlite_graph.py      # SQLite + FTS5 (default)
|   +-- neo4j_graph.py       # Neo4j (production scale)
+-- ingestion/
    |-- metadata.py          # LLM metadata extraction
    |-- pipeline.py          # Single-document ingestion
    +-- batch.py             # Batch ingestion (100K+ docs)

---

Storage Backends

Local Filesystem (Default)

# Automatic — just use Forest() and it stores in ./pageindex_data/
forest = Forest()

Google Cloud Storage

pip install agentic-rag-core[gcs]

from agenticrag import Forest
from agenticrag.storage import GCSStore

store = GCSStore(
    bucket_name="my-bucket",
    prefix="trees/",
    credentials="path/to/service-account.json",
)
forest = Forest(store=store)

Neo4j Graph (Production)

pip install agentic-rag-core[neo4j]

from agenticrag import Forest
from agenticrag.graph import Neo4jGraph

graph = Neo4jGraph(
    uri="bolt://localhost:7687",
    user="neo4j",
    password="your_password",
)
forest = Forest(graph=graph)

---

Tips and Best Practices

Tip	Why
Start with Forest()	Works out of the box, zero config
Use verbose=True	See exactly what the AI agents are doing
Use batch ingestion for 100+ docs	forest.add_directory_batch() is 2-4x faster
Set skip_description=True in batch	Halves the number of LLM calls
Use resume=True in batch	Safely restart interrupted runs
Use skip_critic=True for speed	Faster answers (but less hallucination protection)
Try Qwen3 4B for local LLMs	Best quality-per-VRAM model available (2.5 GB)

---

Hardware Guide (Local LLMs)

Setup	GPU	VRAM	Recommended Model	Speed
Minimum	Any NVIDIA (CUDA >= 5.0)	4-5 GB	Qwen3 4B	Good
Good	RTX 3060 / Quadro P2000	5-12 GB	Qwen3 4B or 8B	Better
Recommended	RTX 4070 Ti	16 GB	Qwen3 8B or 14B	Fast
Ideal	RTX 4090	24 GB	Qwen3 30B	Fastest

Note: AMD GPUs with Polaris/GCN architecture (like RX 580) are not supported by Ollama. Only NVIDIA GPUs with CUDA >= 5.0 and AMD RDNA (RX 5000+) work.

---

FAQ

What file types are supported?

PDF, Markdown (.md), and plain text (.txt) files.

Do I need a GPU?

No. If you use Groq Cloud (free API), everything runs in the cloud. You only need a GPU if you want to run local LLMs via Ollama.

How is this different from LangChain / LlamaIndex?

Traditional RAG (LangChain, LlamaIndex) splits documents into chunks and uses vector similarity to find relevant pieces. This breaks down on professional documents because similarity is not the same as relevance.

AgenticRAG builds a hierarchical tree index and uses LLM reasoning to navigate it — like a human expert flipping through a report. No vectors, no embeddings, no chunking.

How do I get a Groq API key?

1. Go to console.groq.com
2. Sign up (free)
3. Go to API Keys > Create API Key
4. Copy the key that starts with gsk_

Can I use OpenAI / Anthropic / other providers?

AgenticRAG works with any OpenAI-compatible API. Set the base_url parameter to point to your provider's endpoint.

Do I need a `.env` file to run the Web UI?

No. The .env file is completely optional for the Web UI. If you start the server using python -m agenticrag serve without a .env file, the server will start up normally. You can configure your API keys (for Groq or Gemini) or local LLM connection settings directly from the Settings panel in the browser. These will be saved locally to your workspace.

---

License

MIT — free to use in personal and commercial projects.

---

Credits

Architecture inspired by VectifyAI/PageIndex.

Built by Arham Mirkar.