Skip to content

v0.9.0-beta

Pre-release
Pre-release

Choose a tag to compare

View all tags
@vignesh-nagarajan-vn vignesh-nagarajan-vn released this 05 Aug 04:01
v0.9.0-beta
7da47e2

Quarry-LDR turns a research topic into a cited markdown report. The design rests on one idea: the local GPU is a compression layer, not a brain. It squeezes roughly 750K tokens of raw scraped web text into roughly 60K tokens of deduplicated, reranked evidence, and Claude is called only on that dense payload for planning, gap analysis, and synthesis.

Approach Cost per report
All API (~750K raw tokens through Opus) $10 to $15
Quarry-LDR hybrid (~60K compressed tokens to the API) $1.36 to $2.88, measured

The measured numbers come from live validation runs on one laptop card, an NVIDIA RTX 5060 Mobile with 8 GB of VRAM. First Run Report breaks down where every cent went.

Highlights

  • Fourteen checkpointed stages from plan to render; an interrupted run resumes from its last completed stage instead of restarting.
  • Every claim in a report carries a citation that resolves to a source URL and chunk offsets.
  • Three GPU models (embedder, reranker, llama.cpp triage) share one 8 GB card under a VRAM arbiter with a hard 6.5 GB budget.
  • API spend lands in a cost ledger computed from the API's own usage blocks, never estimated from text length. Prompt caching keeps the synthesis corpus warm across the research loop.
  • SearXNG metasearch and a polite fetcher: robots.txt respected, rate limited, cached.
  • Windows, Linux, and WSL2 all work.

Getting Started

An NVIDIA GPU (8 GB VRAM or more), Docker for SearXNG, and an Anthropic API key are required; make bootstrap installs uv and the dependencies.

git clone https://github.com/vignesh-nagarajan-vn/Quarry-LDR
cd Quarry-LDR
make bootstrap                # fresh Windows without GNU make: powershell -ExecutionPolicy Bypass -File scripts/bootstrap.ps1
cp .env.example .env          # paste your ANTHROPIC_API_KEY
make searxng                  # starts local search in Docker
uv run python scripts/download_models.py   # fetches llama-server and the triage GGUF
uv run quarry verify          # preflight check with remediation hints
uv run quarry research "your topic"

Reports land in data/reports/ with a cost ledger and a run manifest. make smoke runs a capped end-to-end rehearsal with a $2.00 ceiling if you want proof before spending on a real topic.

Beta Status and Roadmap

This prerelease is the complete hybrid architecture, validated end to end; it is marked beta because the next major version changes the architecture, not because this one is unfinished. Work on main now targets v1.0.0: a local-first rebuild that moves planning, gap analysis, and synthesis onto local models, makes API calls optional, and adds a branded PDF alongside the markdown report. The hybrid version in this release is preserved on the archive/v0-hybrid-api branch and under this tag, and it remains the right choice when you want Claude writing the report.

Documentation

  • Architecture: pipeline diagram, VRAM arbiter rules, full configuration reference.
  • First Run Report: the live validation story, measured costs, and the bugs only production could find.
  • Example Report: a real report produced during validation for $2.88 of API spend, verbatim except punctuation.
  • Troubleshooting: symptoms, causes, exact fixes.

Known Limitations

  • An Anthropic API key and live spend are required; this version has no offline mode. Planning and synthesis default to Opus, gap checks to Sonnet.
  • NVIDIA CUDA only.
  • Developed and tested on 8 GB cards (RTX 5060 Mobile, desktop RTX 4060). More VRAM is fine; smaller cards are untested.

License: MIT.

Assets 2
Loading