.env.example

# ------------------------------------------------------------------------------
# Host mounts (Docker)
# ------------------------------------------------------------------------------
# Optional: host folder mounted at /workspaces (read-only) inside the container.
# Use this if you want static codebases in config/codebases.yaml like /workspaces/my-repo.
HOST_CODEBASES_ROOT=C:/path/to/projects

# Host filesystem root mounted at /host (read-only). This enables dynamic indexing
# of repos outside /workspaces via MCP register_codebase (example: /host/Users/...).
HOST_FILESYSTEM_ROOT=C:/

# Dockerfile used by mcp-server build.
# Use "Dockerfile" for default image, or "Dockerfile.rocm" for AMD ROCm hosts.
MCP_DOCKERFILE=Dockerfile

# ------------------------------------------------------------------------------
# Models and retrieval pipeline
# ------------------------------------------------------------------------------
# Dense embedding model used for semantic code retrieval.
EMBEDDING_MODEL=jinaai/jina-embeddings-v2-base-code

# Sparse encoder model used for hybrid dense+sparse retrieval.
SPARSE_MODEL=naver/splade-cocondenser-ensembledistil

# Enable sparse branch in retrieval (recommended for recall).
# For very large first-time indexing, set false to speed up ingestion.
USE_SPARSE=true

# Cross-encoder reranker model used after retrieval candidates are fetched.
RERANKER_MODEL=BAAI/bge-reranker-v2-m3

# Enable reranking stage (recommended for precision).
USE_RERANKER=true

# Candidate pool sizing before reranking:
# final_candidates = clamp(limit * multiplier, min, max)
RERANK_CANDIDATE_MULTIPLIER=8
RERANK_MIN_CANDIDATES=30
RERANK_MAX_CANDIDATES=150

# ------------------------------------------------------------------------------
# Devices and batch sizes
# ------------------------------------------------------------------------------
# Device: auto, cuda, or cpu.
EMBEDDING_DEVICE=auto
SPARSE_DEVICE=auto
RERANK_DEVICE=auto

# Throughput tuning (increase with more VRAM/RAM).
EMBEDDING_BATCH_SIZE=32
SPARSE_BATCH_SIZE=16
RERANK_BATCH_SIZE=16

# If true, recreate Qdrant collection when vector schema/dimension mismatches.
# Warning: this can drop existing indexed vectors in that collection.
RECREATE_COLLECTION_ON_DIM_MISMATCH=true

# If true, each Qdrant write waits for persistence/ack and increases indexing latency.
# For faster bulk indexing, keep false.
QDRANT_WRITE_WAIT=false

# ------------------------------------------------------------------------------
# Runtime persistence
# ------------------------------------------------------------------------------
# File used to persist codebases registered at runtime via MCP.
RUNTIME_CODEBASES_FILE=/models/runtime_codebases.yaml

# Directory used to persist incremental indexing state per codebase.
INDEX_STATE_DIR=/models/index_state

# ------------------------------------------------------------------------------
# Search behavior tuning
# ------------------------------------------------------------------------------
# Light score multipliers applied by file type after retrieval/rerank.
CODE_EXTENSION_BOOST=1.08
DOC_EXTENSION_BOOST=0.90

# Default mode: auto, code_only, mixed, docs_only.
SEARCH_MODE_DEFAULT=auto

# Max chunks kept per file before reranking (dedupe control).
MAX_CANDIDATES_PER_FILE=6

# Overlap ratio threshold to treat two chunks as duplicates.
DEDUPE_LINE_OVERLAP_THRESHOLD=0.65

# Character-range overlap threshold used before line-based overlap in dedupe.
# Helps avoid collapsing many chunks from single-line/minified files.
DEDUPE_CHAR_OVERLAP_THRESHOLD=0.65

# ------------------------------------------------------------------------------
# MCP server
# ------------------------------------------------------------------------------
# MCP HTTP port (used both inside container and host publish).
MCP_PORT=8000
MCP_TRANSPORT=streamable-http
MCP_HTTP_PATH=/mcp

# If true, server accepts stateless MCP HTTP calls (no session ID required).
# Enable this if your MCP client reports "Missing session ID".
MCP_STATELESS_HTTP=false

# Async indexing job executor:
# keep as 1 for predictable resource usage; increase if you want queued jobs to run in parallel.
INDEX_JOB_MAX_WORKERS=1

# Max number of recent indexing jobs kept in memory for status/history lookup.
INDEX_JOB_MAX_HISTORY=200

# ------------------------------------------------------------------------------
# Indexing behavior
# ------------------------------------------------------------------------------
# If true, indexes configured codebases on startup.
# Startup indexing runs in background (non-blocking), so MCP endpoint can accept requests immediately.
AUTO_INDEX_ON_STARTUP=true

# If true, startup indexing will force full rebuild instead of incremental.
AUTO_INDEX_ON_STARTUP_FULL_REINDEX=false

# Enable tree-sitter chunking with textual fallback.
USE_TREE_SITTER=true

# Number of chunks buffered across multiple files before an embedding/upsert flush.
# Higher values reduce roundtrips and improve throughput on large repos.
INDEX_CHUNK_BUFFER_SIZE=512

# Qdrant upsert batch size per request.
QDRANT_UPSERT_BATCH_SIZE=256

# Print indexing progress every N scanned files (0 disables progress logs).
INDEX_PROGRESS_EVERY_FILES=250

# Parallel workers for file read/hash/chunk preparation.
# Keep embedding/upsert centralized; this only parallelizes per-file preprocessing.
INDEX_FILE_WORKERS=4

# Max queued file tasks before applying backpressure to the scanner loop.
INDEX_MAX_PENDING_FUTURES=16