# llama.cpp exposes an OpenAI-compatible API when started with llama-server.
LLAMA_CPP_BASE_URL=http://localhost:8080/v1
LLAMA_CPP_MODEL=local-model

# OpenAI-compatible clients still expect a key. llama.cpp does not validate it
# by default, so any non-empty value is fine unless you configure auth.
LLAMA_CPP_API_KEY=not-needed

# Useful while studying deterministic examples.
LLAMA_CPP_TEMPERATURE=0

# Separate llama.cpp embedding server, managed by `llme`.
LLAMA_CPP_EMBED_BASE_URL=http://localhost:8082/v1
# Use "auto" to ask the live embedding server which model is loaded.
LLAMA_CPP_EMBED_MODEL=auto
LLAMA_CPP_EMBED_API_KEY=not-needed

# ---- Embedding provider ----
# Pick one: llama_cpp (default), openai_compat, ai_studio. llama_cpp keeps the
# legacy LLAMA_CPP_EMBED_* defaults above. openai_compat requires
# EMBEDDING_BASE_URL and EMBEDDING_MODEL.
EMBEDDING_PROVIDER=llama_cpp
EMBEDDING_BASE_URL=
EMBEDDING_MODEL=
EMBEDDING_API_KEY=

# AI Studio embeddings use the official google-genai SDK.
# API key fallback order includes AI_STUDIO_EMBED_API_KEY, EMBEDDING_API_KEY,
# AI_STUDIO_API_KEY, GEMINI_API_KEY, and GOOGLE_API_KEY.
AI_STUDIO_EMBED_MODEL=gemini-embedding-2
AI_STUDIO_EMBED_API_KEY=
# For gemini-embedding-2 these choose prompt prefixes; for gemini-embedding-001
# they are sent as the Google task_type.
AI_STUDIO_EMBED_DOCUMENT_TASK_TYPE=RETRIEVAL_DOCUMENT
AI_STUDIO_EMBED_QUERY_TASK_TYPE=RETRIEVAL_QUERY
AI_STUDIO_EMBED_OUTPUT_DIMENSIONALITY=

# Leave unset for defaults. For non-llama providers, the Qdrant vector name is
# auto-derived from provider/model/dimension. Set to "auto" to force the same
# behavior for llama/openai-compatible providers, or set a concrete vector name.
# QDRANT_VECTOR_NAME=auto

# Local Qdrant vector database.
QDRANT_URL=http://localhost:6333
QDRANT_COLLECTION=study_docs
QDRANT_DISTANCE=COSINE

# Chunking defaults for text/markdown ingestion.
RAG_CHUNK_CHARS=900
RAG_CHUNK_OVERLAP=150

# AI-writable knowledge base. Tools are constrained to this root.
KB_ROOT=knowledge_base
KB_MAX_FILE_CHARS=100000
KB_ALLOWED_SUFFIXES=.md,.txt

# summarize_webpage refuses to call the model when extracted page text is larger
# than this budget. Rough heuristic: 4 chars ~= 1 token, so 48000 chars is ~12K
# tokens before prompt overhead in a 16K context.
PAGE_SUMMARY_MAX_CHARS=48000

# Context monitoring/compaction defaults. The agent compacts around 75% so a
# 16K llama.cpp context has room for the summary request and the next response.
CONTEXT_WINDOW_TOKENS=16000
CONTEXT_COMPACT_TRIGGER_RATIO=0.75
CONTEXT_CHARS_PER_TOKEN=4
CONTEXT_SUMMARY_TARGET_TOKENS=1200

# ---- Multi-vendor LLM provider ----
# Pick one: llama_cpp (default), openai, anthropic, ollama, groq, openrouter,
# ai_studio, openai_compat. PROVIDER_MODEL/PROVIDER_TEMPERATURE override the
# vendor default. PROVIDER_BASE_URL overrides the vendor's standard URL and is
# required when PROVIDER=openai_compat.
PROVIDER=llama_cpp
PROVIDER_MODEL=
PROVIDER_TEMPERATURE=0
PROVIDER_BASE_URL=
PROVIDER_API_KEY=
# Optional headers, comma-separated key=value pairs (e.g. for OpenRouter).
PROVIDER_EXTRA_HEADERS=

# Vendor-specific API keys. Only set the ones you actually use.
OPENAI_API_KEY=
ANTHROPIC_API_KEY=
GROQ_API_KEY=
OPENROUTER_API_KEY=
AI_STUDIO_API_KEY=
OLLAMA_BASE_URL=http://localhost:11434/v1
OLLAMA_API_KEY=ollama

# AI Studio-specific Gemini/Gemma behavior when PROVIDER=ai_studio.
# Options: none (default), minimal, low, medium, high.
# Gemini 2.5 Flash can disable thinking with none/budget=0; Gemini 3 can only
# reduce it. Gemma 4 can still emit empty or visible thought-channel tags, so
# output stripping stays enabled by default.
AI_STUDIO_THINKING_LEVEL=none
AI_STUDIO_THINKING_BUDGET=
AI_STUDIO_INCLUDE_THOUGHTS=false
AI_STUDIO_STRIP_THOUGHTS=true

# ---- FastAPI server ----
API_HOST=0.0.0.0
API_PORT=8000
# Comma-separated list, or "*" to allow any origin.
API_CORS_ORIGINS=*
# If set, all /v1/* routes require Authorization: Bearer <token> or X-API-Key.
API_AUTH_TOKEN=
# Set to "true" to make the FastAPI server use the mock graph (no LLM calls).
LANGGRAPH_SERVER_MOCK=false