.env.example

# Wilson Environment Configuration
# Copy this file to .env and fill in your values
# Never commit .env to git

# CourtListener API Token
# Get yours free at https://www.courtlistener.com/sign-in/
COURTLISTENER_TOKEN=your_courtlistener_token_here

# PACER Credentials (optional — enables original filing retrieval)
# Register free at https://pacer.uscourts.gov
PACER_USERNAME=your_pacer_username_here
PACER_PASSWORD=your_pacer_password_here

# Local LLM Configuration (optional -- enables Phase 3 coherence checking)
# Wilson uses Ollama for local inference. Install at https://ollama.com
#
# Recommended models (pull with: ollama pull <model>):
#
#   qwen3.5:35b          -- Best balance of quality and speed. MoE architecture
#                           runs fast even at 35B. Supports 128k context.
#                           Requires ~22GB VRAM. Recommended for RTX 3090/4090/5090.
#
#   nemotron-cascade-2:30b -- NVIDIA MoE model, strong legal reasoning.
#                             Requires ~23GB VRAM. 128k context.
#
#   qwen2.5:14b          -- Good quality, fits in 10GB VRAM. 128k context.
#                           Recommended for mid-range GPUs.
#
#   llama3.1:8b          -- Minimum viable. 5GB VRAM. 128k context.
#                           Use if VRAM is limited.
#
#   llama3:latest        -- NOT recommended. 32k context causes truncation
#                           into dissent sections, producing wrong verdicts.
#
# Do NOT use models with less than 64k context for coherence checking.
# Short context causes the model to read the dissent instead of the holding.
OLLAMA_HOST=http://localhost:11434
OLLAMA_MODEL=qwen3.5:35b

# Context window size in tokens.
# Larger = more opinion text analyzed = better verdicts.
# Wilson truncates opinions to fit; front-loaded to preserve the holding.
#
# Hardware guide:
#   32GB VRAM (RTX 5090): 245760 (240k) with MoE models (qwen3.5:35b, nemotron-cascade-2:30b)
#   24GB VRAM (RTX 4090): 65536 (64k) safe for most models
#   16GB VRAM:            32768 (32k) -- upgrade model if verdict quality is poor
#   CPU-only:             8192  (8k)  -- slow, limited quality
#
OLLAMA_CONTEXT_SIZE=245760

# Wilson Data Paths
# Where Wilson stores bulk data downloads
# Defaults to ~/wilson-data if not set
WILSON_DATA_PATH=~/wilson-data

# Path to CourtListener bulk citations CSV (optional — enables offline verification)
# Download from CourtListener bulk data:
# https://com-courtlistener-storage.s3-us-west-2.amazonaws.com/bulk-data/
# File: citations-YYYY-MM-DD.csv.bz2 (~121MB compressed, ~1.9GB uncompressed)
# Decompress with: bunzip2 citations-YYYY-MM-DD.csv.bz2
CITATIONS_CSV=/path/to/citations-2026-03-31.csv

# Charlotin database path (optional — enables batch processing)
# Download from https://www.damiencharlotin.com/hallucinations/
CHARLOTIN_CSV=data/Charlotin-hallucination_cases.csv
RESULTS_CSV=data/charlotin_results.csv