Advanced RAG System - Persian/Arabic Support

📤 Upload Documents

📁 Drag & Drop files here or click to browse

Supported: PDF, DOCX, TXT, Images (PNG, JPG, TIFF)

Max files per upload: loading…

🗂️ Target Collection

🧠 Embedding Model Model used to convert documents into vectors. Must match the collection's vector dimension.

🔍 Enable OCR (for scanned documents and images)

Chunking Strategy

Sentence Window — splits into overlapping chunks (articles, manuals, books) Whole Document — no splitting, one vector per file (poems, short texts) Semantic — split by meaning (embedding-based; good for mixed-topic docs)

Processing documents...

💬 Ask Questions

🗂️ Collection

🧠 Embedding Model

🔀 Reranker Model

✏️ Your question

Searching...

Your answers will appear here...

📊 System Statistics

0

Documents

0

Queries

Not Ready

Index Status

detecting…

GPU

🕵️ Recent Query Audit (last 5)

Per-phase timings help find bottlenecks. Slow retrieve → reduce num_queries, disable multi-query, or lower vector_top_k. Slow rewrite → disable query rewriting. Slow synthesize → LLM latency.

No queries recorded yet.

⚙️ RAG Settings

🔑 LLM Provider & API Keys

LLM Provider: .env default: —

OpenAI API Key

Anthropic API Key

🔌 Custom LLM API (OpenAI-like / Yarabot / Rakhshai)

API URL

Model / Engine

API Key (if required)

Choose provider above, then fill the matching API fields. Rakhshai: URL https://rakhshai.ir:2083/v1, engine iran-llm-zal, Bearer API key (docs). Yarabot: full completion URL + model id. OpenAI-like: base URL ending in /v1.

Vector top_k

BM25 top_k

Fusion top_k

Num queries (RRF)

Rerank top_n

Memory length Previous Q&A turns passed to LLM (0 = off)

Sentence window Surrounding sentences per chunk (re-upload to apply)

System prompt to LLM

📤 Upload Limits

Controls the dashboard upload tab. .env defaults: — files, — MB per file.

Max files per upload Single request limit (1–500)

Max file size (MB) Per-file size limit (1–1000 MB)

💬 Open WebUI Integration

Control how adv-rag handles chat UI requests. adv-rag owns prompts, retrieval, and source display — Open WebUI should not inject its own RAG or translation into your queries.

Strip translation instructions — remove [OUTPUT_LANGUAGE: …] appended by chat UI

Strip Open WebUI RAG injection — ignore <source> / template context in user messages

Use chat history from Open WebUI — sanitized previous turns for conversation memory

Append sources markdown to answer — add **Sources:** block at end of response text

Show sources in Open WebUI citation panel — native sources UI below the answer (instead of markdown block)

Source snippet length (chars)

Reranker Model

Download from HuggingFace

Use downloaded model

Cross-encoder / reranker models only — not for embeddings.

Embedding Settings Changing provider or model requires re-uploading all documents (vectors become incompatible).

Provider:

Model:

.env defaults — Provider: - | Model: -

🔧 Query Engineering

Enable Query Rewriting - Rewrite queries to be more specific for better retrieval

Enable Multi-Query Generation - Generate multiple query variations for broader coverage

Enable HyDE (Hypothetical Document Embeddings) - Generate a hypothetical answer and use it for retrieval (improves recall)

Query Rewrite Prompt Template

Multi-Query Prompt Template

🚀 Advanced RAG System