# Copy to .env and fill in. All variables are optional; defaults match deploy/docker-compose.yml.
# Read by bce.config.Settings (prefix BCE_). .env is gitignored.

# --- PostgreSQL (single DB: AGE + pgvector + SQL) ---
BCE_DB_HOST=localhost
BCE_DB_PORT=5432
BCE_DB_NAME=bce
BCE_DB_USER=bce
BCE_DB_PASSWORD=bce
# Name of the Apache AGE graph holding the code graph.
BCE_GRAPH_NAME=code_graph

# --- i18n (presentation only; never affects payload) ---
BCE_DEFAULT_LOCALE=en
# Locales the API and UI accept, as a JSON array. Anything else falls back to the default.
BCE_SUPPORTED_LOCALES=["en","tr"]

# --- CORS ---
# Browser origins allowed to call the API, as a JSON array. The bundled UI is served from the
# API's own origin and needs no entry here; this only matters for a separately hosted frontend
# such as the Vite dev server. Use ["*"] for development only.
BCE_CORS_ORIGINS=["http://localhost:5173","http://127.0.0.1:5173"]

# --- Embedding (Phase 2; anchor finding only, pinned for determinism P2) ---
# Provider: "hashing" (default, dependency-free deterministic fallback), "voyage" (Voyage AI) or
# "openai" (any OpenAI-compatible /v1/embeddings server: vLLM, TEI, Ollama, OpenAI - this is how
# an open model such as jinaai/jina-code-embeddings-1.5b runs on-prem).
BCE_EMBEDDING_PROVIDER=hashing
# Model name + dimension are pinned together as the stored model_id; changing either is a reindex.
# `bce migrate` re-types the pgvector column to BCE_EMBEDDING_DIM (refusing to drop stored vectors
# unless run with --reset-embeddings).
# Supported (see README "Supported models"); the retrieval profiles were fitted on these two:
#   voyage: BCE_EMBEDDING_MODEL=voyage-code-4              BCE_EMBEDDING_DIM=1024
#   openai: BCE_EMBEDDING_MODEL=jina-code-embeddings-1.5b  BCE_EMBEDDING_DIM=1536
# (voyage-code-3, 1024, still works and stays the code default for indexes built with it.)
BCE_EMBEDDING_MODEL=voyage-code-4
BCE_EMBEDDING_DIM=1024
# Required when BCE_EMBEDDING_PROVIDER=voyage, which also needs the optional 'voyageai' package:
# pip install "bgts-context-engine[embed]". Selecting voyage without both is an error, not a
# silent fall back to hashing -- index and query vectors have to come from the same encoder.
BCE_VOYAGE_API_KEY=
# --- provider "openai" only ---
# e.g. vllm serve ~/models/jina-code-embeddings-1.5b --runner pooling --port 8001 \
#        --served-model-name jina-code-embeddings-1.5b
BCE_EMBEDDING_BASE_URL=http://127.0.0.1:8001/v1
# Bearer token; leave empty for a local server.
BCE_EMBEDDING_API_KEY=
# Instruction prefixes for instruction-tuned models. Unset = built-in default from the model name
# (jina-code-* gets its nl2code query/passage prompts; anything else none). Set to override.
# BCE_EMBEDDING_QUERY_PREFIX=
# BCE_EMBEDDING_DOCUMENT_PREFIX=
# Extra JSON merged into every request. The default makes vLLM truncate inputs longer than its
# --max-model-len instead of rejecting the batch. OpenAI's own API rejects unknown fields: use {}.
BCE_EMBEDDING_EXTRA_BODY={"truncate_prompt_tokens": -1}
# Seconds per embedding request while indexing (a local model can take minutes on a large batch).
BCE_EMBEDDING_TIMEOUT=600
# Seconds per *query* embedding at search time (one short text; two attempts). Kept short so a tool
# call returns - without its semantic anchor, lexical/structural anchors still answer - before an
# agent's own MCP timeout (Cursor: 60 s) kills it.
BCE_EMBEDDING_QUERY_TIMEOUT=10
# --- Retrieval profile: engine constants that depend on the embedding model ---
# "auto" picks the profile fitted for BCE_EMBEDDING_MODEL: voyage-* -> "voyage" (the historical
# constants), jina-* -> "jina" (model's top-10 ranks guarded in narrowing), anything else -> the
# voyage set. Name one to force it. The BCE_RETRIEVAL_SEMANTIC_* knobs override single values of the
# chosen profile and exist for fitting runs (local_bench/bench40_jina.py rerun); leave them unset.
BCE_RETRIEVAL_PROFILE=auto
# BCE_RETRIEVAL_SEMANTIC_GUARD_RANKS=
# BCE_RETRIEVAL_SEMANTIC_RESERVE_SHARE=
# BCE_RETRIEVAL_SEMANTIC_RANK_SCALE=

# --- Context selector (optional; see docs/selector.md) ---
# A decision model that tiers the ranked answer after retrieval: the files the task most likely
# edits in full, likely related files as one line each, the rest dropped. Values:
#   off         (default) no selector: get_context_for_task returns the plain ranking at K=20
#   jev         TypeSafe's Jev, hosted (OpenRouter or TypeSafe's own API)
#   decider-2b  open-weight https://huggingface.co/Mapika/decider-2b on your own GPU
#   decider-4b  open-weight https://huggingface.co/Mapika/decider-4b on your own GPU
# With a selector set, the server ranks K=50 and the model cuts it down. It is fail-open: an
# error or timeout returns the plain ranking, and coverage.selector says so.
BCE_SELECTOR=off
#
# Example 1 - Jev over OpenRouter (no TypeSafe account; billed to the OpenRouter account):
#   BCE_SELECTOR=jev
#   OPENROUTER_API_KEY=sk-or-v1-...
#   (defaults: BCE_SELECTOR_MODEL=typesafe/jev-1.13
#              BCE_SELECTOR_URL=https://openrouter.ai/api/alpha/decisions  BCE_SELECTOR_TIMEOUT=3)
#
# Example 2 - Jev on TypeSafe's own API (BCE_SELECTOR_API_KEY wins over the OpenRouter key):
#   BCE_SELECTOR=jev
#   BCE_SELECTOR_URL=https://api.typesafe.ai/v1/systemone
#   BCE_SELECTOR_API_KEY=<TYPESAFE_API_KEY>
#   BCE_SELECTOR_MODEL=jev-latest
#
# Example 3 - decider-4b (or decider-2b) behind `uvicorn decider.serve:app` on a GPU host, e.g. a
# RunPod TCP port mapping. No key; the value must match the model the server loaded.
#   BCE_SELECTOR=decider-4b
#   BCE_SELECTOR_URL=http://<host>:<port>/v1/systemone
#   (defaults: BCE_SELECTOR_URL=http://127.0.0.1:8000/v1/systemone  BCE_SELECTOR_TIMEOUT=10)
#
# OpenRouter key for Jev. Set either name, not both: an empty BCE_OPENROUTER_API_KEY= line in this
# file hides an OPENROUTER_API_KEY below it.
# OPENROUTER_API_KEY=
# Bearer token for the selector endpoint: TypeSafe's key, or an auth proxy's in front of a decider.
# BCE_SELECTOR_API_KEY=
# Empty = the default of the chosen selector (see above).
BCE_SELECTOR_MODEL=
BCE_SELECTOR_URL=
# Seconds per request, also the worst-case latency the selector adds. Unset = 3 (jev) / 10 (decider).
# BCE_SELECTOR_TIMEOUT=
# Tiering policy, fitted on Jev and shared by every model. Leave unset unless re-fitting.
# BCE_SELECTOR_FULL_THRESHOLD=0.5
# BCE_SELECTOR_STUB_THRESHOLD=0.05
# BCE_SELECTOR_MAX_FILES=12
# BCE_SELECTOR_SYMBOL_MIN_SCORE=1.0

# --- Bitbucket Cloud remote indexing (bce index-remote / POST /v1/index-remote) ---
# Access token (recommended): leave username empty; we clone as x-token-auth:<token>.
# App password: also set BCE_BITBUCKET_USERNAME to your Bitbucket username.
BCE_BITBUCKET_USERNAME=
BCE_BITBUCKET_TOKEN=
# Where cloned repos are cached (re-index does a fast git fetch). Gitignored.
BCE_REPO_CACHE_DIR=.bce_data/repos
# Set to false ONLY behind a TLS-intercepting corporate proxy.
BCE_GIT_SSL_VERIFY=true

# --- Logging (JSON lines to stdout; optional rotating file sink) ---
# DEBUG logs request start, body, response and every pipeline stage; INFO logs summaries only.
BCE_LOG_LEVEL=INFO
# Set to true to ALSO write logs to a rotating file (path below).
BCE_LOG_FILE_ENABLED=false
BCE_LOG_FILE_PATH=.bce_data/logs/bce.log
BCE_LOG_FILE_MAX_BYTES=10485760
BCE_LOG_FILE_BACKUP_COUNT=5

# --- Background jobs (async indexing via POST /v1/jobs/* and the MCP indexing tools) ---
# Worker threads running inside the API server process (1 = jobs run strictly in order).
BCE_JOB_WORKERS=1
# Seconds a worker sleeps between queue polls when idle.
BCE_JOB_POLL_INTERVAL=2.0

# --- MCP (bce serve-mcp) ---
# Off: the server advertises read-only tools only. On: it also exposes index_repo and
# reindex_repo (local working trees; cloning stays on REST/CLI), and runs its own job workers so
# queued indexing progresses without a separate `bce serve`.
BCE_MCP_ALLOW_WRITE=false
# Comma-separated tool names the MCP server advertises; empty = all read tools. Agents fetch a
# tool's schema before first use and chain small calls when many tools are offered, so a narrow
# catalog costs fewer model turns. The agent benchmark ran with the single primary tool:
#   BCE_MCP_TOOLS=get_context_for_task
BCE_MCP_TOOLS=
# How the agent uses the get_context_for_task answer; the server states the workflow in the tool
# description and in every answer's payload.workflow. hint (empty = default): the answer is the
# starting point - add the files of uncovered identifiers, search only when the engine says the
# answer is likely incomplete. trust: the answer is the set of locations - no searching the tree.
# The init commands write it into the editor's MCP config, which wins over this file.
BCE_AGENT_MODE=

# Not a setting, but worth knowing: BCE_ENV_FILE (or `bce --env-file PATH`) points the engine at an
# explicit .env. Editors spawn `bce serve-mcp` from their own working directory, so without it the
# MCP server never reads this file. `bce --env-file PATH cursor-init` / `claude-init` write the
# editor config (and the agent rule) for a project with that path filled in.