Akshara Vision

Restore. Read. Preserve.

GitHub CLI ↗

What is Akshara?

A local-first, vendor-agnostic document intelligence CLI & grounded chat workbench. It turns unstructured book scans and historical archives into normalized layout trees and publication-ready digital volumes.

The Problem Space

Vision language models and LLMs are excellent at OCR correction, but integrating them directly into bulk archival pipelines fails on practical execution.

Ad-Hoc Model CallsAkshara Engine
Fragile PipelinesAPI failures or exits abort progress; no state preservation.Resumable RunsPage checkpoints guarantee zero loss on interruption.
Raw Text ClutterExtracts plain strings without layout tags or semantic roles.Structure ParsingBuilds normalized trees mapping columns, headers, and crops.
Vendor Lock-inHardcoded API connections restrict backend migration.Agnostic RoutingUniform routing across local GGUF engines or cloud APIs.

Layout Normalization

Akshara Vision does not treat scans as plain text blocks. Every page is parsed hierarchically, extracting reading orders, columns, and semantic document roles.

Semantic Units

Identifies preface, table of contents, chapter boundaries, footnotes, running headers, and page numbers.

Normalized Layout Trees

Reconstructs page columns and flow. Captures exact native layout bounding boxes, page zones, and confidence indexes.

Figure Alignment

Identifies visual figures and tables, generates conservative crop bounding boxes, and automatically captions visual elements.

Grounded CLI Review & Chat

Query processed run directories directly using indexed source chunks for strict audit trails. Audit layouts, confidence scores, and crop regions natively.

Core Capabilities

Model Routing

Transparent query routing across local offline GGUF engines (Ollama, llama.cpp) for zero cost, or cloud APIs (Gemini, Sarvam, Anthropic).

Fault-Tolerant Runs

Writes page and chunk level stage checkpoints. Interruptions or crashes do not force re-processing; resume picks up exactly where it left off.

Structure Cache

Stores raw structural layout trees. Segregates running headers, tables, images, and footers from reading flow to compile clean source copies.

Rich Exporters

Compiles structured books into multiple publication formats. Supports plain text, standard Markdown, EPUB packages, and formatted PDF printouts.

Infrastructure & Stack

Local GGUF Engines (Ollama, llama.cpp)
Cloud Vision APIs (Gemini, Sarvam)
Layout Parsers (doctr, paddle, layoutparser)
Engine Checkpointing (SQLite)
Document Exporters (EPUB, DOCX, styled PDF)