Legal Query

Full JSON Response

📰 News Collector

Search Bangladeshi news sources in English or Bangla. Keywords are auto-translated for cross-language search.

Keywords

From Date

To Date

📄 Analysis Report

📂 Collected Articles (0)

📋 Raw JSON Data

Recent Sessions

📚 Batch Index

Scan a local directory of PDFs and index them in parallel. Each file is dispatched as a separate task with live progress monitoring. Already-indexed files (by content SHA-256) are skipped.

Previous Batch Runs

📄 Documents

Manage indexed documents. Upload new PDFs, view document list, or delete documents. Uploaded files are indexed with the same pipeline (OCR, Bangla normalization, chunking, embedding) as batch-indexed files.

Upload Document

PDF or text file

Title (optional)

Indexed Documents

⚙ Tasks

View the task queue: indexing, news collection, KG updates, and scanner runs. Tasks are processed by the in-process worker pool.

Task Queue

Auto-refresh (3s)

⚙ Configuration

Settings are persisted to .data/rag_config.yaml and loaded on every restart. Changes apply immediately (model singletons reload on save).

LLM Connection Provider

Provider type

Model name / path

Server URL / API base

API key (optional for local)

GPU layers (llama.cpp)

Temperature

Max tokens

Embedding Models

Dense text backend

Text model name

Sparse text backend

Image embedder

Image model name / GGUF path

Image server URL (llama-server only)

Embedder API key (Cohere/OpenAI)

GPU Acceleration

Backend

Device id

Min free VRAM (MB)

VRAM budget (MB, 0=auto 90%)

CPU fallback on low VRAM

Memory manager serializes local-GPU model calls and runs remote-API calls in parallel, preventing VRAM OOM on mixed workloads.

Hybrid Retrieval

Fusion strategy

Dense weight (α, for weighted)

RRF k (for rrf)

Final top-k

Query Processing

Query rewriting

Max rewrite variants

Confidence high threshold

⚠ Reset all data?

This will permanently erase:

All indexed documents, chunks, and embeddings (SQLite DB)
The in-memory vector store + BM25 index
The AI semantic cache
Raw uploaded files (PDFs, images, HTML) under .data/objects
All loaded model singletons (embedder, reranker, LLM provider, retriever, query processor)

This cannot be undone. After reset, restart the server to re-bootstrap the admin user, then re-index your documents.