Settings reference

Generated by scripts/gen_settings_doc.py from src/lasagna/core/config.py and src/lasagna/settings.lasagna.toml. Do not edit by hand: run the script.

Where a setting comes from, lowest priority first:

1. the default below;
2. settings.lasagna.toml, [rag] table (LASAGNA_SETTINGS_FILE names another file);
3. the environment variable: LASAGNA_<NAME>, or RAG_<NAME> (the older name, still read; LASAGNA_ wins when both are set);
4. the session (lasagna configure --profile <name> -s <session>, the REST API or MCP).

The settings file and environment variables are the process defaults. A session's profile and overrides are kept in the session database and win over both for that session only.

An environment variable that does not parse, or is outside the allowed values, stops the process with an error naming it, except log_level, which warns and keeps its default: it is also the lasagna command's console level (WARNING unless set), and a typo in it must not stop lasagna doctor. A RAG_ variable that is not a setting is ignored with a warning that suggests the closest one. Booleans take true/false, 1/0, yes/no or on/off; lists and objects take JSON.

Data lives in ~/.lasagna (LASAGNA_DATA_DIR): the databases in dbs/ (LASAGNA_DB_DIR) and uploads in uploads/ (LASAGNA_UPLOAD_DIR). A ~/.rag_knowledge from an earlier version is moved there on first use.

Vector Database

Setting	Type	Default	Environment variables	Description
vector_backend	str	"milvus"	LASAGNA_VECTOR_BACKEND, RAG_VECTOR_BACKEND	Vector database backend to use.
vector_backend_config	object	{}	LASAGNA_VECTOR_BACKEND_CONFIG, RAG_VECTOR_BACKEND_CONFIG	Backend-specific configuration.
collection_name	str	"rag_documents_v2"	LASAGNA_COLLECTION_NAME, RAG_COLLECTION_NAME	The vector collection this configuration reads and writes.
embedding_dim	int, 1 to 16384	1024	LASAGNA_EMBEDDING_DIM, RAG_EMBEDDING_DIM	Vector size; an embedding_model set without one brings its own size.
metric_type	str	"COSINE"	LASAGNA_METRIC_TYPE, RAG_METRIC_TYPE	vector similarity: COSINE, L2 or IP.
index_type	str	"HNSW"	LASAGNA_INDEX_TYPE, RAG_INDEX_TYPE	dense index on servers; Milvus Lite always uses FLAT.
index_params	object	{"M": 16, "efConstruction": 256}	LASAGNA_INDEX_PARAMS, RAG_INDEX_PARAMS	HNSW build parameters.
search_params	object	{"ef": 128}	LASAGNA_SEARCH_PARAMS, RAG_SEARCH_PARAMS	HNSW search breadth.

Embedding Model

Setting	Type	Default	Environment variables	Description
embedding_provider	str	"huggingface"	LASAGNA_EMBEDDING_PROVIDER, RAG_EMBEDDING_PROVIDER	where embeddings are computed.
embedding_model	str	"BAAI/bge-m3"	LASAGNA_EMBEDDING_MODEL, RAG_EMBEDDING_MODEL	dense+sparse+ColBERT, 8K context, 100+ languages.
embedding_model_fallback	str	"BAAI/bge-large-en-v1.5"	LASAGNA_EMBEDDING_MODEL_FALLBACK, RAG_EMBEDDING_MODEL_FALLBACK	loaded if the model fails; never during evaluation.
embedding_backend	str	"auto"	LASAGNA_EMBEDDING_BACKEND, RAG_EMBEDDING_BACKEND	Embedding backend: 'auto', 'sentence-transformers', or 'flag-embedding'.
normalize_embeddings	bool	true	LASAGNA_NORMALIZE_EMBEDDINGS, RAG_NORMALIZE_EMBEDDINGS	Normalize embedding vectors for cosine similarity. Required for most models.
batch_size	int	32	LASAGNA_BATCH_SIZE, RAG_BATCH_SIZE	How many text pieces to process at once. Higher = faster but uses more memory.
use_gpu	bool	true	LASAGNA_USE_GPU, RAG_USE_GPU	Process documents faster using your graphics card, if available.
embedding_truncate_dim	int or none	none	LASAGNA_EMBEDDING_TRUNCATE_DIM, RAG_EMBEDDING_TRUNCATE_DIM	Matryoshka dimension truncation. None = use full dim.

Chunking

Setting	Type	Default	Environment variables	Description
chunking_strategy	str	"hierarchical"	LASAGNA_CHUNKING_STRATEGY, RAG_CHUNKING_STRATEGY	what runs: enable_hierarchical, then preserve_headers or preserve_tables, override another name.
chunk_size	int, 1 to 100000	800	LASAGNA_CHUNK_SIZE, RAG_CHUNK_SIZE	How large each piece of content should be (in characters). Smaller = more precise, larger = more context.
chunk_overlap	int, 0 to 100000	0	LASAGNA_CHUNK_OVERLAP, RAG_CHUNK_OVERLAP	Research: no measurable benefit from overlap (arXiv:2601.14123)
semantic_threshold	float	0.75	LASAGNA_SEMANTIC_THRESHOLD, RAG_SEMANTIC_THRESHOLD	How sensitive the smart splitter is to topic changes. Lower = more splits, higher = fewer splits.
semantic_breakpoint_method	str	"absolute"	LASAGNA_SEMANTIC_BREAKPOINT_METHOD, RAG_SEMANTIC_BREAKPOINT_METHOD	How the semantic chunker picks breakpoints: 'absolute' or a Kamradt method.
semantic_breakpoint_amount	float or none	none	LASAGNA_SEMANTIC_BREAKPOINT_AMOUNT, RAG_SEMANTIC_BREAKPOINT_AMOUNT	The breakpoint method's parameter; None uses the method's default.
semantic_buffer_size	int	1	LASAGNA_SEMANTIC_BUFFER_SIZE, RAG_SEMANTIC_BUFFER_SIZE	Neighbouring sentences embedded with each sentence, on either side.
min_chunk_size	int	200	LASAGNA_MIN_CHUNK_SIZE, RAG_MIN_CHUNK_SIZE	Pieces smaller than this are merged with their neighbors.
max_chunk_size	int	1500	LASAGNA_MAX_CHUNK_SIZE, RAG_MAX_CHUNK_SIZE	Pieces larger than this are force-split to keep them manageable.
window_size	int	3	LASAGNA_WINDOW_SIZE, RAG_WINDOW_SIZE	Number of sentences per chunk in sentence_window strategy.
sentence_overlap	int	1	LASAGNA_SENTENCE_OVERLAP, RAG_SENTENCE_OVERLAP	Number of sentences to overlap in sentence_window strategy.
enable_hierarchical	bool	true	LASAGNA_ENABLE_HIERARCHICAL, RAG_ENABLE_HIERARCHICAL	Keeps related pieces of a document together so answers include full context, not just fragments.
parent_chunk_size	int	2000	LASAGNA_PARENT_CHUNK_SIZE, RAG_PARENT_CHUNK_SIZE	Size of the broader context pieces (in characters). Provides surrounding context when needed.
child_chunk_size	int	500	LASAGNA_CHILD_CHUNK_SIZE, RAG_CHILD_CHUNK_SIZE	Size of the precise matching pieces (in characters). Used for initial search.
auto_merge_threshold	float	0.5	LASAGNA_AUTO_MERGE_THRESHOLD, RAG_AUTO_MERGE_THRESHOLD	When enough small pieces match from the same section, automatically use the full section instead.
enable_late_chunking	bool	false	LASAGNA_ENABLE_LATE_CHUNKING, RAG_ENABLE_LATE_CHUNKING	Late chunking: embed full doc, pool per chunk (Jina, arxiv:2409.04701)
late_chunking_overlap	int	512	LASAGNA_LATE_CHUNKING_OVERLAP, RAG_LATE_CHUNKING_OVERLAP	Tokens shared by consecutive windows in long late chunking.
chunk_unit	str	"chars"	LASAGNA_CHUNK_UNIT, RAG_CHUNK_UNIT	What chunk_size and chunk_overlap count: 'chars', 'words' or 'tokens'.
chunk_tokenizer	str or none	none	LASAGNA_CHUNK_TOKENIZER, RAG_CHUNK_TOKENIZER	Tokenizer for chunk_unit='tokens': a Hugging Face name, or 'tiktoken:'.
recursive_separators	list	["\n\n", "\n", ". ", "? ", "! ", " ", ""]	LASAGNA_RECURSIVE_SEPARATORS, RAG_RECURSIVE_SEPARATORS	Separators the recursive chunker tries, coarsest first.
recursive_keep_separator	str	"end"	LASAGNA_RECURSIVE_KEEP_SEPARATOR, RAG_RECURSIVE_KEEP_SEPARATOR	Whether a separator ends the piece before it ('end') or starts the next ('start').
markdown_header_levels	int	6	LASAGNA_MARKDOWN_HEADER_LEVELS, RAG_MARKDOWN_HEADER_LEVELS	Deepest Markdown header level the markdown_header chunker splits on (1-6).
markdown_strip_headers	bool	true	LASAGNA_MARKDOWN_STRIP_HEADERS, RAG_MARKDOWN_STRIP_HEADERS	Leave header lines out of markdown_header chunk text (they are in metadata).
markdown_split_long_sections	bool	true	LASAGNA_MARKDOWN_SPLIT_LONG_SECTIONS, RAG_MARKDOWN_SPLIT_LONG_SECTIONS	Split Markdown sections longer than chunk_size with the recursive chunker.
proposition_model	str or none	none	LASAGNA_PROPOSITION_MODEL, RAG_PROPOSITION_MODEL	LLM that writes propositions; None uses the configured default.
proposition_passage_size	int	100	LASAGNA_PROPOSITION_PASSAGE_SIZE, RAG_PROPOSITION_PASSAGE_SIZE	Words per passage the proposition chunker hands to the LLM.
contextual_base_chunker	str	"recursive"	LASAGNA_CONTEXTUAL_BASE_CHUNKER, RAG_CONTEXTUAL_BASE_CHUNKER	Registered chunker whose chunks the contextual chunker situates.
contextual_model	str or none	none	LASAGNA_CONTEXTUAL_MODEL, RAG_CONTEXTUAL_MODEL	LLM that writes each chunk's context; None uses the configured default.
contextual_max_document_chars	int or none	50000	LASAGNA_CONTEXTUAL_MAX_DOCUMENT_CHARS, RAG_CONTEXTUAL_MAX_DOCUMENT_CHARS	Characters of the document in each context prompt; None for all of it.
preserve_tables	bool	true	LASAGNA_PRESERVE_TABLES, RAG_PRESERVE_TABLES	Prevents tables from being split across multiple pieces.
preserve_headers	bool	true	LASAGNA_PRESERVE_HEADERS, RAG_PRESERVE_HEADERS	Attaches section headings to their content so context is never lost.
extract_metadata	bool	true	LASAGNA_EXTRACT_METADATA, RAG_EXTRACT_METADATA	keep page, section and file metadata on each chunk.
max_file_size	int	52428800	LASAGNA_MAX_FILE_SIZE, RAG_MAX_FILE_SIZE	50MB.
max_content_length	int	4000	LASAGNA_MAX_CONTENT_LENGTH, RAG_MAX_CONTENT_LENGTH	Maximum content length per chunk in storage (characters).

Hybrid Search

Setting	Type	Default	Environment variables	Description
enable_hybrid_search	bool	true	LASAGNA_ENABLE_HYBRID_SEARCH, RAG_ENABLE_HYBRID_SEARCH	Finds results by both meaning and exact keywords, so it catches acronyms, IDs, and rare terms that pure AI search misses.
hybrid_alpha	float, 0.0 to 1.0	0.7	LASAGNA_HYBRID_ALPHA, RAG_HYBRID_ALPHA	Balance between understanding meaning (right) and matching exact words (left). Most documents work best around 0.7.
sparse_method	str	"bm25"	LASAGNA_SPARSE_METHOD, RAG_SPARSE_METHOD	Which keyword-based retrieval method to use alongside AI vector search.
default_top_k	int, 1 to 10000	20	LASAGNA_DEFAULT_TOP_K, RAG_DEFAULT_TOP_K	Research: 75+ introduces hard negatives (ICLR 2025)
similarity_threshold	float, 0.0 to 1.0	0.3	LASAGNA_SIMILARITY_THRESHOLD, RAG_SIMILARITY_THRESHOLD	Cut applied to the per-query normalized score when the agentic loop or an evaluation's generation track searches. Searches from the CLI, REST, MCP and the SDK apply no threshold unless the request sets min_score.
rrf_k	int	60	LASAGNA_RRF_K, RAG_RRF_K	RRF constant for Reciprocal Rank Fusion.
enable_bm25_stemming	bool	true	LASAGNA_ENABLE_BM25_STEMMING, RAG_ENABLE_BM25_STEMMING	Enable Porter stemming for BM25.
bm25_max_sessions	int	1000	LASAGNA_BM25_MAX_SESSIONS, RAG_BM25_MAX_SESSIONS	Maximum number of in-memory BM25 session indices to keep.
bm25_session_ttl_seconds	int	86400	LASAGNA_BM25_SESSION_TTL_SECONDS, RAG_BM25_SESSION_TTL_SECONDS	Idle TTL for BM25 session indices in seconds.
bm25_k1	float	1.2	LASAGNA_BM25_K1, RAG_BM25_K1	BM25 term-frequency saturation for the bm25 and hybrid retrievers.
bm25_b	float	0.75	LASAGNA_BM25_B, RAG_BM25_B	BM25 document-length normalisation for the bm25 and hybrid retrievers.
splade_model	str	"naver/splade-cocondenser-ensembledistil"	LASAGNA_SPLADE_MODEL, RAG_SPLADE_MODEL	Masked-language-model checkpoint the splade retriever encodes with.
hybrid_fusion_depth	int	1000	LASAGNA_HYBRID_FUSION_DEPTH, RAG_HYBRID_FUSION_DEPTH	Candidates each list contributes to hybrid_rrf / hybrid_convex fusion.
hybrid_convex_normalization	str	"tmm"	LASAGNA_HYBRID_CONVEX_NORMALIZATION, RAG_HYBRID_CONVEX_NORMALIZATION	Score normalisation for hybrid_convex: 'tmm', 'minmax' or 'zscore'.

Reranking

Setting	Type	Default	Environment variables	Description
enable_reranking	bool	true	LASAGNA_ENABLE_RERANKING, RAG_ENABLE_RERANKING	Double-checks results with a precision model to put the most relevant ones first.
reranker_model	str	"mixedbread-ai/mxbai-rerank-large-v1"	LASAGNA_RERANKER_MODEL, RAG_RERANKER_MODEL	BEIR NDCG@10 = 57.49.
rerank_top_k	int, 1 to 1000	5	LASAGNA_RERANK_TOP_K, RAG_RERANK_TOP_K	Research: context cliff at ~2,500 tokens.
reranker_batch_size	int	16	LASAGNA_RERANKER_BATCH_SIZE, RAG_RERANKER_BATCH_SIZE	How many results to score at once. Higher = faster but uses more memory.
custom_reranker	str	""	LASAGNA_CUSTOM_RERANKER, RAG_CUSTOM_RERANKER	Registry name for a custom reranker component (empty string = use default).
custom_query_enhancer	str	""	LASAGNA_CUSTOM_QUERY_ENHANCER, RAG_CUSTOM_QUERY_ENHANCER	Registry name for a custom query enhancer component (empty string = use default).
llm_reranker_model	str	""	LASAGNA_LLM_RERANKER_MODEL, RAG_LLM_RERANKER_MODEL	Model for LLM listwise reranker (empty config = configured default).
llm_reranker_window_size	int	20	LASAGNA_LLM_RERANKER_WINDOW_SIZE, RAG_LLM_RERANKER_WINDOW_SIZE	Sliding window size for listwise reranking.
llm_reranker_step_size	int	10	LASAGNA_LLM_RERANKER_STEP_SIZE, RAG_LLM_RERANKER_STEP_SIZE	How far the listwise reranker's window moves toward the top each step.
colbert_model	str	"colbert-ir/colbertv2.0"	LASAGNA_COLBERT_MODEL, RAG_COLBERT_MODEL	ColBERT late-interaction model name.
monot5_model	str	"castorini/monot5-base-msmarco-10k"	LASAGNA_MONOT5_MODEL, RAG_MONOT5_MODEL	Sequence-to-sequence checkpoint the monot5 reranker scores with.
flashrank_model	str	"ms-marco-TinyBERT-L-2-v2"	LASAGNA_FLASHRANK_MODEL, RAG_FLASHRANK_MODEL	FlashRank model name for the flashrank reranker.
cohere_rerank_model	str	"rerank-v3.5"	LASAGNA_COHERE_RERANK_MODEL, RAG_COHERE_RERANK_MODEL	Cohere Rerank model for the cohere reranker.
reranker_trust_remote_code	bool	false	LASAGNA_RERANKER_TRUST_REMOTE_CODE, RAG_RERANKER_TRUST_REMOTE_CODE	Let the cross-encoder run modelling code shipped with its checkpoint.
enable_context_reordering	bool	true	LASAGNA_ENABLE_CONTEXT_REORDERING, RAG_ENABLE_CONTEXT_REORDERING	Lost-in-the-middle fix: reorder results so best are at start and end.
enable_context_compression	bool	false	LASAGNA_ENABLE_CONTEXT_COMPRESSION, RAG_ENABLE_CONTEXT_COMPRESSION	Enable context compression to reduce noise/tokens after reranking.
compression_method	str	"extractive"	LASAGNA_COMPRESSION_METHOD, RAG_COMPRESSION_METHOD	Compression method: 'extractive' or 'abstractive'.
compression_ratio	float	0.3	LASAGNA_COMPRESSION_RATIO, RAG_COMPRESSION_RATIO	Target compression ratio (0.3 = keep 30%).
enable_graph_retrieval	bool	false	LASAGNA_ENABLE_GRAPH_RETRIEVAL, RAG_ENABLE_GRAPH_RETRIEVAL	Enable GraphRAG entity-linked retrieval.
graph_max_hops	int	2	LASAGNA_GRAPH_MAX_HOPS, RAG_GRAPH_MAX_HOPS	Max traversal depth in knowledge graph.
graph_entity_types	list	["PERSON", "ORG", "CONCEPT", "LOCATION", "TECHNOLOGY", "E...	LASAGNA_GRAPH_ENTITY_TYPES, RAG_GRAPH_ENTITY_TYPES	Entity types for graph extraction.
graph_extraction_model	str	""	LASAGNA_GRAPH_EXTRACTION_MODEL, RAG_GRAPH_EXTRACTION_MODEL	Model for entity extraction (empty config = configured default).
agentic_prompts	object	{}	LASAGNA_AGENTIC_PROMPTS, RAG_AGENTIC_PROMPTS	Per-node prompt overrides for the agentic graph, keyed by node name.
agentic_models	object	{}	LASAGNA_AGENTIC_MODELS, RAG_AGENTIC_MODELS	Per-node model overrides for the agentic graph, keyed by node name.
enable_raptor	bool	false	LASAGNA_ENABLE_RAPTOR, RAG_ENABLE_RAPTOR	Enable RAPTOR recursive summary tree indexing.
raptor_max_cluster_size	int	10	LASAGNA_RAPTOR_MAX_CLUSTER_SIZE, RAG_RAPTOR_MAX_CLUSTER_SIZE	Max chunks per cluster in RAPTOR tree.
raptor_max_levels	int	5	LASAGNA_RAPTOR_MAX_LEVELS, RAG_RAPTOR_MAX_LEVELS	Max tree depth.
enable_speculative_rag	bool	false	LASAGNA_ENABLE_SPECULATIVE_RAG, RAG_ENABLE_SPECULATIVE_RAG	Enable speculative RAG: parallel draft generation + verification.
speculative_n_drafts	int	3	LASAGNA_SPECULATIVE_N_DRAFTS, RAG_SPECULATIVE_N_DRAFTS	Number of parallel draft answers for speculative RAG.
speculative_subset_size	int	3	LASAGNA_SPECULATIVE_SUBSET_SIZE, RAG_SPECULATIVE_SUBSET_SIZE	Number of context chunks per draft in speculative RAG.
self_ask_max_follow_ups	int	5	LASAGNA_SELF_ASK_MAX_FOLLOW_UPS, RAG_SELF_ASK_MAX_FOLLOW_UPS	Self-Ask: follow-up questions searched before the model must conclude.
ircot_paragraphs_per_step	int	6	LASAGNA_IRCOT_PARAGRAPHS_PER_STEP, RAG_IRCOT_PARAGRAPHS_PER_STEP	IRCoT: paragraphs retrieved per step (K).
ircot_max_steps	int	8	LASAGNA_IRCOT_MAX_STEPS, RAG_IRCOT_MAX_STEPS	IRCoT: reasoning steps before the loop stops.
ircot_max_paragraphs	int	15	LASAGNA_IRCOT_MAX_PARAGRAPHS, RAG_IRCOT_MAX_PARAGRAPHS	IRCoT: cap on the paragraphs collected across steps.
flare_retrieval_threshold	float	0.8	LASAGNA_FLARE_RETRIEVAL_THRESHOLD, RAG_FLARE_RETRIEVAL_THRESHOLD	FLARE theta: retrieve when any token of the next sentence is below it.
flare_mask_threshold	float	0.4	LASAGNA_FLARE_MASK_THRESHOLD, RAG_FLARE_MASK_THRESHOLD	FLARE beta: tokens below it are masked out of the retrieval query.
flare_max_sentences	int	8	LASAGNA_FLARE_MAX_SENTENCES, RAG_FLARE_MAX_SENTENCES	FLARE: sentences generated before the loop stops.
flare_docs_per_retrieval	int	2	LASAGNA_FLARE_DOCS_PER_RETRIEVAL, RAG_FLARE_DOCS_PER_RETRIEVAL	FLARE: documents retrieved per retrieval.
flare_confidence_source	str	"auto"	LASAGNA_FLARE_CONFIDENCE_SOURCE, RAG_FLARE_CONFIDENCE_SOURCE	FLARE: 'auto', 'logprobs' or 'self_reported' confidence.
adaptive_rag_classifier_model	str	""	LASAGNA_ADAPTIVE_RAG_CLASSIFIER_MODEL, RAG_ADAPTIVE_RAG_CLASSIFIER_MODEL	Adaptive-RAG: model for the complexity classifier (empty = the loop's).
adaptive_rag_single_step_docs	int	5	LASAGNA_ADAPTIVE_RAG_SINGLE_STEP_DOCS, RAG_ADAPTIVE_RAG_SINGLE_STEP_DOCS	Adaptive-RAG: documents retrieved on the single-step route.
self_rag_passages	int	5	LASAGNA_SELF_RAG_PASSAGES, RAG_SELF_RAG_PASSAGES	Self-RAG: passages retrieved, each answered from and critiqued.
self_rag_weight_rel	float	1.0	LASAGNA_SELF_RAG_WEIGHT_REL, RAG_SELF_RAG_WEIGHT_REL	Self-RAG: weight of the relevance critique (ISREL).
self_rag_weight_sup	float	1.0	LASAGNA_SELF_RAG_WEIGHT_SUP, RAG_SELF_RAG_WEIGHT_SUP	Self-RAG: weight of the support critique (ISSUP).
self_rag_weight_use	float	0.5	LASAGNA_SELF_RAG_WEIGHT_USE, RAG_SELF_RAG_WEIGHT_USE	Self-RAG: weight of the utility critique (ISUSE).
self_rag_retrieval	str	"adaptive"	LASAGNA_SELF_RAG_RETRIEVAL, RAG_SELF_RAG_RETRIEVAL	Self-RAG: 'adaptive' (the model judges) or 'always' retrieve.
enable_mmr	bool	true	LASAGNA_ENABLE_MMR, RAG_ENABLE_MMR	Avoids showing repetitive or near-duplicate results, so answers draw on varied sources.
mmr_lambda	float, 0.0 to 1.0	0.5	LASAGNA_MMR_LAMBDA, RAG_MMR_LAMBDA	LangChain and LlamaIndex default; a pre-MMR filter keeps only relevant items.
mmr_score_normalization	str	"sigmoid"	LASAGNA_MMR_SCORE_NORMALIZATION, RAG_MMR_SCORE_NORMALIZATION	MMR score normalization strategy.

Query Enhancement

Setting	Type	Default	Environment variables	Description
enable_hyde	bool	false	LASAGNA_ENABLE_HYDE, RAG_ENABLE_HYDE	Uses AI to imagine what the ideal answer looks like, then searches for content matching that answer.
hyde_model	str	"default"	LASAGNA_HYDE_MODEL, RAG_HYDE_MODEL	Which AI model powers the AI-Enhanced Search feature. 'default' uses your main model.
hyde_n_docs	int	3	LASAGNA_HYDE_N_DOCS, RAG_HYDE_N_DOCS	Number of hypothetical documents for multi-HyDE.
hyde_embedding_strategy	str	"average"	LASAGNA_HYDE_EMBEDDING_STRATEGY, RAG_HYDE_EMBEDDING_STRATEGY	Strategy for combining multi-HyDE embeddings.
enable_query_expansion	bool	false	LASAGNA_ENABLE_QUERY_EXPANSION, RAG_ENABLE_QUERY_EXPANSION	Automatically rephrase your question in different ways to find more results.
max_query_variations	int	3	LASAGNA_MAX_QUERY_VARIATIONS, RAG_MAX_QUERY_VARIATIONS	How many alternative phrasings to try when Search Variations is enabled.
enable_query_decomposition	bool	false	LASAGNA_ENABLE_QUERY_DECOMPOSITION, RAG_ENABLE_QUERY_DECOMPOSITION	Split multi-part questions into simpler sub-questions and search each separately.
enable_query_classification	bool	true	LASAGNA_ENABLE_QUERY_CLASSIFICATION, RAG_ENABLE_QUERY_CLASSIFICATION	Enable adaptive query-type classification.
agentic_hyde_strategy	str	"adaptive"	LASAGNA_AGENTIC_HYDE_STRATEGY, RAG_AGENTIC_HYDE_STRATEGY	When to use AI-Enhanced Search during multi-step searches.
enable_contextual_retrieval	bool	false	LASAGNA_ENABLE_CONTEXTUAL_RETRIEVAL, RAG_ENABLE_CONTEXTUAL_RETRIEVAL	Enable Anthropic-style contextual retrieval.
enable_rag_fusion	bool	false	LASAGNA_ENABLE_RAG_FUSION, RAG_ENABLE_RAG_FUSION	Enable RAG Fusion: multi-query generation + Reciprocal Rank Fusion.
rag_fusion_n_queries	int	3	LASAGNA_RAG_FUSION_N_QUERIES, RAG_RAG_FUSION_N_QUERIES	Number of query variations to generate for RAG Fusion.
query2doc_format	str	"dense"	LASAGNA_QUERY2DOC_FORMAT, RAG_QUERY2DOC_FORMAT	Query2doc: 'dense' (q [SEP] d') or 'sparse' (q repeated, then d').
query2doc_query_repeats	int	5	LASAGNA_QUERY2DOC_QUERY_REPEATS, RAG_QUERY2DOC_QUERY_REPEATS	Query2doc: how many times the query precedes its pseudo-document (sparse form).
query2doc_max_tokens	int	128	LASAGNA_QUERY2DOC_MAX_TOKENS, RAG_QUERY2DOC_MAX_TOKENS	Query2doc: token budget for the generated pseudo-document.
enable_crag	bool	false	LASAGNA_ENABLE_CRAG, RAG_ENABLE_CRAG	Enable Corrective RAG: LLM grades retrieval quality post-search.
crag_web_search_fallback	bool	false	LASAGNA_CRAG_WEB_SEARCH_FALLBACK, RAG_CRAG_WEB_SEARCH_FALLBACK	Enable web search fallback for Corrective RAG.

Query Routing (Adaptive-RAG)

Setting	Type	Default	Environment variables	Description
enable_query_routing	bool	true	LASAGNA_ENABLE_QUERY_ROUTING, RAG_ENABLE_QUERY_ROUTING	Route each query to the fast, standard or comprehensive tier by its estimated complexity.
routing_strategy	str: rule_based, adaptive	"rule_based"	LASAGNA_ROUTING_STRATEGY, RAG_ROUTING_STRATEGY	Routing strategy: 'rule_based' (deterministic) or 'adaptive' (confidence-aware).
routing_default_tier	str: fast, standard, comprehensive	"standard"	LASAGNA_ROUTING_DEFAULT_TIER, RAG_ROUTING_DEFAULT_TIER	Default routing tier when routing is disabled.
routing_confidence_threshold	float, 0.0 to 1.0	0.5	LASAGNA_ROUTING_CONFIDENCE_THRESHOLD, RAG_ROUTING_CONFIDENCE_THRESHOLD	Confidence threshold below which the adaptive strategy escalates the tier.
routing_allow_tier_override	bool	true	LASAGNA_ROUTING_ALLOW_TIER_OVERRIDE, RAG_ROUTING_ALLOW_TIER_OVERRIDE	Allow API callers to force a specific routing tier.
routing_log_decisions	bool	true	LASAGNA_ROUTING_LOG_DECISIONS, RAG_ROUTING_LOG_DECISIONS	Log routing decisions for observability.

Citation Tracking

Setting	Type	Default	Environment variables	Description
citation_format	str	"stable_id"	LASAGNA_CITATION_FORMAT, RAG_CITATION_FORMAT	How citations are identified.
citation_id_length	int	8	LASAGNA_CITATION_ID_LENGTH, RAG_CITATION_ID_LENGTH	Length of stable citation IDs (when using stable_id format).
verify_citations	bool	true	LASAGNA_VERIFY_CITATIONS, RAG_VERIFY_CITATIONS	Verify that cited chunks actually exist.
track_citations_db	bool	true	LASAGNA_TRACK_CITATIONS_DB, RAG_TRACK_CITATIONS_DB	Persist citation history to the database.
show_page_numbers	bool	true	LASAGNA_SHOW_PAGE_NUMBERS, RAG_SHOW_PAGE_NUMBERS	Include page numbers in citations (when available).
include_source_preview	bool	true	LASAGNA_INCLUDE_SOURCE_PREVIEW, RAG_INCLUDE_SOURCE_PREVIEW	a short excerpt of the source with each citation.
citation_db_max_age_hours	int	168	LASAGNA_CITATION_DB_MAX_AGE_HOURS, RAG_CITATION_DB_MAX_AGE_HOURS	1 week.
citation_db_max_per_session	int	1000	LASAGNA_CITATION_DB_MAX_PER_SESSION, RAG_CITATION_DB_MAX_PER_SESSION	oldest citations beyond this are removed.
citation_normalize_scores	bool	true	LASAGNA_CITATION_NORMALIZE_SCORES, RAG_CITATION_NORMALIZE_SCORES	scale citation scores to 0-1.
citation_score_method	str	"minmax"	LASAGNA_CITATION_SCORE_METHOD, RAG_CITATION_SCORE_METHOD	'minmax', 'sigmoid', 'clamp'.

File Types

Setting	Type	Default	Environment variables	Description
supported_extensions	list	[".pdf", ".docx", ".txt", ".md", ".csv", ".json", ".html"...	LASAGNA_SUPPORTED_EXTENSIONS, RAG_SUPPORTED_EXTENSIONS	Only formats DocumentProcessor can identify and extract. .doc, .rtf and .epub have no extractor, and .pptx and .xlsx are refused by the upload type check before Docling could convert them, so listing them advertised uploads that always failed.

Storage

Setting	Type	Default	Environment variables	Description
upload_dir_name	str	"rag_uploads"	LASAGNA_UPLOAD_DIR_NAME, RAG_UPLOAD_DIR_NAME	folder name for stored uploads.
db_name	str	"rag.db"	LASAGNA_DB_NAME, RAG_DB_NAME	Unified database (consolidated from rag_metadata.db + rag_citations.db)

Parallel Upload

Setting	Type	Default	Environment variables	Description
max_concurrent_uploads	int, 1 to 20	5	LASAGNA_MAX_CONCURRENT_UPLOADS, RAG_MAX_CONCURRENT_UPLOADS	Maximum number of parallel document uploads (1-20).

OCR

Setting	Type	Default	Environment variables	Description
enable_ocr	bool	false	LASAGNA_ENABLE_OCR, RAG_ENABLE_OCR	Read text from scanned PDFs and images. Useful for old documents or photos of text.
ocr_engine	str	"easyocr"	LASAGNA_OCR_ENGINE, RAG_OCR_ENGINE	Which engine reads text from scanned documents and images.
ocr_languages	list	["en"]	LASAGNA_OCR_LANGUAGES, RAG_OCR_LANGUAGES	Language codes for OCR text extraction.
ocr_confidence_threshold	float	0.5	LASAGNA_OCR_CONFIDENCE_THRESHOLD, RAG_OCR_CONFIDENCE_THRESHOLD	How confident the OCR must be before accepting text. Lower = more text (possibly noisy), higher = cleaner.

Processing

Setting	Type	Default	Environment variables	Description
extract_images	bool	false	LASAGNA_EXTRACT_IMAGES, RAG_EXTRACT_IMAGES	extract images from documents.
preserve_formatting	bool	true	LASAGNA_PRESERVE_FORMATTING, RAG_PRESERVE_FORMATTING	keep layout such as lists and line breaks in extracted text.
async_processing	bool	true	LASAGNA_ASYNC_PROCESSING, RAG_ASYNC_PROCESSING	process uploads in the background.
processing_timeout	int	300	LASAGNA_PROCESSING_TIMEOUT, RAG_PROCESSING_TIMEOUT	seconds a document may take to process.
allow_model_download_in_requests	bool	false	LASAGNA_ALLOW_MODEL_DOWNLOAD_IN_REQUESTS, RAG_ALLOW_MODEL_DOWNLOAD_IN_REQUESTS	Whether a server may download a missing model while it handles a request.

Docling Document Processing

Setting	Type	Default	Environment variables	Description
docling_enabled	bool	true	LASAGNA_DOCLING_ENABLED, RAG_DOCLING_ENABLED	Enable Docling for structure-aware document processing.
docling_mode	str	"auto"	LASAGNA_DOCLING_MODE, RAG_DOCLING_MODE	Docling deployment mode: 'auto', 'local', or 'production'.
docling_pdf_backend	str	"docling_parse"	LASAGNA_DOCLING_PDF_BACKEND, RAG_DOCLING_PDF_BACKEND	Docling PDF backend: 'docling_parse' or 'pypdfium2'.
docling_ocr_enabled	bool	true	LASAGNA_DOCLING_OCR_ENABLED, RAG_DOCLING_OCR_ENABLED	Enable OCR within Docling pipeline.
docling_table_structure	bool	true	LASAGNA_DOCLING_TABLE_STRUCTURE, RAG_DOCLING_TABLE_STRUCTURE	Enable TableFormer table structure extraction.
docling_max_tokens	int, 64 to 32768	512	LASAGNA_DOCLING_MAX_TOKENS, RAG_DOCLING_MAX_TOKENS	Max tokens per chunk for Docling HybridChunker.
docling_timeout	int	300	LASAGNA_DOCLING_TIMEOUT, RAG_DOCLING_TIMEOUT	Per-document processing timeout in seconds.

Caching

Setting	Type	Default	Environment variables	Description
enable_embedding_cache	bool	true	LASAGNA_ENABLE_EMBEDDING_CACHE, RAG_ENABLE_EMBEDDING_CACHE	Cache computed embeddings to avoid recomputation.
embedding_cache_size	int	10000	LASAGNA_EMBEDDING_CACHE_SIZE, RAG_EMBEDDING_CACHE_SIZE	Maximum number of cached embeddings.
enable_search_cache	bool	true	LASAGNA_ENABLE_SEARCH_CACHE, RAG_ENABLE_SEARCH_CACHE	Cache search results for repeated queries.
search_cache_ttl	int	300	LASAGNA_SEARCH_CACHE_TTL, RAG_SEARCH_CACHE_TTL	How long cached search results remain valid (seconds).
search_cache_max_size	int	1000	LASAGNA_SEARCH_CACHE_MAX_SIZE, RAG_SEARCH_CACHE_MAX_SIZE	Maximum number of cached search results.

Security Limits

Setting	Type	Default	Environment variables	Description
max_document_ids_filter	int	100	LASAGNA_MAX_DOCUMENT_IDS_FILTER, RAG_MAX_DOCUMENT_IDS_FILTER	Maximum document IDs allowed in filter expressions.

Observability

Setting	Type	Default	Environment variables	Description
enable_prometheus	bool	false	LASAGNA_ENABLE_PROMETHEUS, RAG_ENABLE_PROMETHEUS	Enable Prometheus metrics collection.
enable_tracing	bool	false	LASAGNA_ENABLE_TRACING, RAG_ENABLE_TRACING	Enable OpenTelemetry distributed tracing.
tracing_exporter	str	"otlp"	LASAGNA_TRACING_EXPORTER, RAG_TRACING_EXPORTER	Tracing exporter: 'otlp', 'jaeger', 'zipkin', or 'console'.
tracing_endpoint	str	"http://localhost:4317"	LASAGNA_TRACING_ENDPOINT, RAG_TRACING_ENDPOINT	OTLP/Jaeger/Zipkin collector endpoint.
tracing_service_name	str	"rag-knowledge"	LASAGNA_TRACING_SERVICE_NAME, RAG_TRACING_SERVICE_NAME	Service name reported in traces.
log_level	str: TRACE, DEBUG, INFO, SUCCESS, WARNING, ERROR, CRITICAL	"INFO"	LASAGNA_LOG_LEVEL, RAG_LOG_LEVEL	Logging verbosity.

Multi-Tenancy

Setting	Type	Default	Environment variables	Description
tenant_isolation	str: none, partition	"none"	LASAGNA_TENANT_ISOLATION, RAG_TENANT_ISOLATION	Tenant isolation mode: 'none' or 'partition'.
quota_enabled	bool	false	LASAGNA_QUOTA_ENABLED, RAG_QUOTA_ENABLED	Enable per-tenant resource quotas (documents, storage, rate limits).

Rate Limiting

Setting	Type	Default	Environment variables	Description
rate_limiting_enabled	bool	false	LASAGNA_RATE_LIMITING_ENABLED, RAG_RATE_LIMITING_ENABLED	Enable adaptive rate limiting per tenant.
rate_limiting_strategy	str	"adaptive"	LASAGNA_RATE_LIMITING_STRATEGY, RAG_RATE_LIMITING_STRATEGY	Rate limiting strategy: 'adaptive' or 'fixed'.
rate_limiting_rpm	float	60.0	LASAGNA_RATE_LIMITING_RPM, RAG_RATE_LIMITING_RPM	Base requests per minute per tenant.
rate_limiting_burst	float	10.0	LASAGNA_RATE_LIMITING_BURST, RAG_RATE_LIMITING_BURST	Maximum burst token capacity per tenant.
rate_limiting_burst_multiplier	float	2.0	LASAGNA_RATE_LIMITING_BURST_MULTIPLIER, RAG_RATE_LIMITING_BURST_MULTIPLIER	Capacity multiplier when system is idle (allows bursting).
rate_limiting_overload_threshold	float	0.8	LASAGNA_RATE_LIMITING_OVERLOAD_THRESHOLD, RAG_RATE_LIMITING_OVERLOAD_THRESHOLD	System load score that triggers throttling (0.0 to 1.0).
rate_limiting_idle_threshold	float	0.3	LASAGNA_RATE_LIMITING_IDLE_THRESHOLD, RAG_RATE_LIMITING_IDLE_THRESHOLD	System load score below which bursting is allowed (0.0 to 1.0).

Database Backend

Setting	Type	Default	Environment variables	Description
database_backend	str: sqlite, postgresql	"sqlite"	LASAGNA_DATABASE_BACKEND, RAG_DATABASE_BACKEND	Database backend: 'sqlite' (default) or 'postgresql'.
database_dsn	str	""	LASAGNA_DATABASE_DSN, RAG_DATABASE_DSN	PostgreSQL connection string (DSN). Required when database_backend='postgresql'.
database_pool_min	int	2	LASAGNA_DATABASE_POOL_MIN, RAG_DATABASE_POOL_MIN	Minimum database connection pool size (PostgreSQL).
database_pool_max	int	10	LASAGNA_DATABASE_POOL_MAX, RAG_DATABASE_POOL_MAX	Maximum database connection pool size (PostgreSQL).

Redis / Distributed Cache

Setting	Type	Default	Environment variables	Description
cache_backend	str: memory, redis	"memory"	LASAGNA_CACHE_BACKEND, RAG_CACHE_BACKEND	Cache backend: 'memory' (default) or 'redis'.
redis_url	str	"redis://localhost:6379"	LASAGNA_REDIS_URL, RAG_REDIS_URL	Redis connection URL.
redis_pool_size	int	10	LASAGNA_REDIS_POOL_SIZE, RAG_REDIS_POOL_SIZE	Redis connection pool size.

Celery / Background Processing

Setting	Type	Default	Environment variables	Description
celery_enabled	bool	false	LASAGNA_CELERY_ENABLED, RAG_CELERY_ENABLED	Enable Celery for background document processing.
celery_broker_url	str	"redis://localhost:6379/0"	LASAGNA_CELERY_BROKER_URL, RAG_CELERY_BROKER_URL	Celery broker URL (typically Redis).
celery_result_backend	str	"redis://localhost:6379/1"	LASAGNA_CELERY_RESULT_BACKEND, RAG_CELERY_RESULT_BACKEND	Celery result backend URL (typically Redis).

Connection Pooling

Setting	Type	Default	Environment variables	Description
connection_pool_size	int, 1 to 500	1	LASAGNA_CONNECTION_POOL_SIZE, RAG_CONNECTION_POOL_SIZE	Vector backend connection pool size. 1 = no pooling (default).

File Storage Backend

Setting	Type	Default	Environment variables	Description
storage_backend	str	"local"	LASAGNA_STORAGE_BACKEND, RAG_STORAGE_BACKEND	File storage backend: 'local' (default), 's3', or 'fsspec'.
storage_backend_config	object	{}	LASAGNA_STORAGE_BACKEND_CONFIG, RAG_STORAGE_BACKEND_CONFIG	Backend-specific storage configuration.

Fast Ingestion (eval-only pipeline tuning)

Setting	Type	Default	Environment variables	Description
ingest_extract_workers	int	min(os.cpu_count() or 4, 8)	LASAGNA_INGEST_EXTRACT_WORKERS, RAG_INGEST_EXTRACT_WORKERS	Number of parallel workers for text extraction/chunking in fast ingest.
ingest_embed_batch_size	int	64	LASAGNA_INGEST_EMBED_BATCH_SIZE, RAG_INGEST_EMBED_BATCH_SIZE	Cross-document embedding batch size for fast ingest.
ingest_insert_batch_size	int	500	LASAGNA_INGEST_INSERT_BATCH_SIZE, RAG_INGEST_INSERT_BATCH_SIZE	Chunks per vector store insert call in fast ingest.

Agentic retrieval and document awareness (settings file only)

Setting	Type	Default	Environment variables	Description
use_agentic_rag	bool	true	LASAGNA_USE_AGENTIC_RAG, RAG_USE_AGENTIC_RAG	Agentic RAG with multi-hop reasoning.
use_final_search	bool	false	LASAGNA_USE_FINAL_SEARCH, RAG_USE_FINAL_SEARCH	Do final search after refinement for complete citations.
max_agent_iterations	int	3	LASAGNA_MAX_AGENT_ITERATIONS, RAG_MAX_AGENT_ITERATIONS	Maximum number of search iterations for agentic RAG.
chunk_refinement_method	str	"auto"	LASAGNA_CHUNK_REFINEMENT_METHOD, RAG_CHUNK_REFINEMENT_METHOD	Whether the AI cleans up and highlights the most relevant parts of found content.
agent_temperature	float	0.1	LASAGNA_AGENT_TEMPERATURE, RAG_AGENT_TEMPERATURE	Temperature for agent/refinement model (low for determinism).
max_query_length	int	12	LASAGNA_MAX_QUERY_LENGTH, RAG_MAX_QUERY_LENGTH	Maximum query length in tokens for refinement queries.
auto_refine_avg_length_threshold	int	1000	LASAGNA_AUTO_REFINE_AVG_LENGTH_THRESHOLD, RAG_AUTO_REFINE_AVG_LENGTH_THRESHOLD	Average chunk length threshold to trigger auto-refinement.
auto_refine_avg_score_threshold	float	0.5	LASAGNA_AUTO_REFINE_AVG_SCORE_THRESHOLD, RAG_AUTO_REFINE_AVG_SCORE_THRESHOLD	Average score threshold below which to trigger auto-refinement.
auto_refine_chunk_count_threshold	int	7	LASAGNA_AUTO_REFINE_CHUNK_COUNT_THRESHOLD, RAG_AUTO_REFINE_CHUNK_COUNT_THRESHOLD	Chunk count threshold above which to trigger auto-refinement.
use_structured_graph	bool	false	LASAGNA_USE_STRUCTURED_GRAPH, RAG_USE_STRUCTURED_GRAPH	Use structured StateGraph instead of flat ReAct agent.
enable_hallucination_check	bool	false	LASAGNA_ENABLE_HALLUCINATION_CHECK, RAG_ENABLE_HALLUCINATION_CHECK	Enable post-generation hallucination grounding check (ISSUP).
enable_answer_quality_check	bool	false	LASAGNA_ENABLE_ANSWER_QUALITY_CHECK, RAG_ENABLE_ANSWER_QUALITY_CHECK	Enable post-generation answer quality check (ISUSE).
max_quality_check_attempts	int, 1 to 20	2	LASAGNA_MAX_QUALITY_CHECK_ATTEMPTS, RAG_MAX_QUALITY_CHECK_ATTEMPTS	Maximum retry attempts for quality gate failures.
coverage_scorer_mode	str	"keyword"	LASAGNA_COVERAGE_SCORER_MODE, RAG_COVERAGE_SCORER_MODE	Coverage assessment mode: 'keyword' (fast) or 'llm' (accurate).
final_search_mode	str	"merge"	LASAGNA_FINAL_SEARCH_MODE, RAG_FINAL_SEARCH_MODE	Final search behavior: 'merge' (combine hop + final) or 'replace' (legacy).
chunk_cite_retrieval	object	{"strategy": "adaptive", "fixed_size": 5, "threshold_scor...	LASAGNA_CHUNK_CITE_RETRIEVAL, RAG_CHUNK_CITE_RETRIEVAL	How many retrieved chunks are kept: a fixed count, or cut at a score gap.
document_awareness	object	{"enabled": true, "enrichment_strategy": "first_tokens", ...	LASAGNA_DOCUMENT_AWARENESS, RAG_DOCUMENT_AWARENESS	What the agent is told about each document: a preview of its first tokens, an LLM summary, or both.
agentic_pipeline	str	none	LASAGNA_AGENTIC_PIPELINE, RAG_AGENTIC_PIPELINE	Name of the registered agentic pipeline to use.
refinement_model	str	none	LASAGNA_REFINEMENT_MODEL, RAG_REFINEMENT_MODEL	Which AI model refines the search results. A fast, affordable model works well here.

Variables that are not settings

The servers, the MCP server and the vector stores read these from the environment only: not from the settings file, and not per session.

REST API and authentication

Variable	Default	Description
RAG_REQUIRE_AUTH	false	When true, only a verified X-Session-Token names a session, and POST /api/v1/sessions needs an X-API-Key. When false, binding beyond loopback is refused at startup unless RAG_BIND_IS_PROTECTED is set.
RAG_API_KEYS	unset	Comma-separated keys, each at least 32 characters, that may mint session tokens when RAG_REQUIRE_AUTH is on. Rate limits and quotas are counted per key, over all the sessions it minted.
RAG_JWT_SECRET	a random secret per process	Signs session tokens (HMAC-SHA256). Unset, tokens stop working when the process restarts; with RAG_REQUIRE_AUTH on, an unset, placeholder or shorter-than-32-character secret stops the server at startup.
RAG_SESSION_TOKEN_TTL_SECONDS	604800	How long a session token is valid, in seconds (7 days).
RAG_ADMIN_TOKEN	unset	The X-Admin-Token for /api/v1/admin/* and for naming a session in POST /api/v1/sessions. Unset, a placeholder or shorter than 32 characters: the admin routes answer 503.
RAG_ALLOW_ANONYMOUS_SESSION	false	With RAG_REQUIRE_AUTH on, keep the anonymous default session usable.
RAG_BIND_IS_PROTECTED	false	Says that something outside the process (a loopback port mapping, a NetworkPolicy, a firewall) restricts access, so the server may bind beyond loopback with auth off.
RAG_ALLOW_HEADER_IDENTITY	false	Accept the X-User-ID header as the user identity; otherwise it is refused with 401.
RAG_CORS_ORIGINS	empty	Comma-separated origins allowed cross-origin requests (REST and MCP over HTTP). Empty allows none; * is refused at startup when RAG_REQUIRE_AUTH is on.
RAG_ALLOWED_HOSTS	unset	Extra Host names the REST API and MCP over HTTP answer to with auth off, comma-separated (name, or name:* for any port), such as a container's service name. Otherwise only localhost, 127.0.0.1 and [::1].
RAG_FORWARDED_ALLOW_IPS	127.0.0.1	Proxy addresses whose X-Forwarded-For is trusted. The image reads gunicorn's FORWARDED_ALLOW_IPS first.
RAG_MAX_BATCH_FILES	20	Files in one POST /api/v1/documents/batch; more is refused with 400.
RAG_MAX_BATCH_BYTES	209715200	Total body of one batch upload (200 MiB); more is refused with 413 before the body is read.
RAG_SEARCH_TIMEOUT	120	Seconds a REST search may run before it answers 504.
RAG_MAX_SESSIONS_PER_KEY	100	With RAG_REQUIRE_AUTH on, how many sessions one API key may mint per token lifetime (RAG_SESSION_TOKEN_TTL_SECONDS); one more answers 429. Raise it for an application that makes a session per end user.
RAG_EMBED_MCP	false	Mount the MCP server at /mcp inside the REST process, for single-process deployments such as Milvus Lite.
RAG_CONFIGURATOR	false	Mount the configurator UI at /configurator, with its API routes.
LASAGNA_LOG_FORMAT	text	json writes one JSON object per log line; text shows a request's id before the message. Any other value stops the server at startup.
PROMETHEUS_MULTIPROC_DIR	unset	The directory gunicorn workers share for Prometheus metrics (set in the image).
RAG_WORKER_METRICS_PORT	unset	The port a Celery worker serves its Prometheus /metrics on, with RAG_ENABLE_PROMETHEUS on (lasagna_ingests_total for the documents it processed). Unset, the worker serves none. Compose and Kubernetes set 9100.

MCP server

Variable	Default	Description
LASAGNA_MCP_TOOLS	safe	safe leaves out the six tools that delete data or reset configuration; all serves every tool. The same as lasagna-mcp --tools.
LASAGNA_ENABLE_DOCS_API	false	Turn on the /api/call tool tester, which calls any tool without authentication. For development only.
RAG_SESSION_ID	unset	The session an MCP tool call, or a REST request without X-Session-ID, uses when it names none (else default). lasagna serve -s sets it.
RAG_MCP_UPLOAD_ROOT	unset	The directory a networked MCP server may ingest server-side files from; unset, it reads none.

Vector store connections

Variable	Default	Description
MILVUS_URI	unset	A Milvus server's address, used when vector_backend_config has no uri. Without one the milvus backend opens Milvus Lite in the data directory.
MILVUS_TOKEN	unset	The token for MILVUS_URI.
QDRANT_URL	unset	A Qdrant server's address; without one the qdrant backend stores its data in the data directory.
QDRANT_API_KEY	unset	The API key for QDRANT_URL.
WEAVIATE_URL	unset	A Weaviate server's address.
WEAVIATE_API_KEY	unset	The API key for WEAVIATE_URL.
PINECONE_API_KEY	unset	The Pinecone API key the pinecone backend needs.
PGVECTOR_DSN	postgresql://rag:rag@localhost:5432/rag	The PostgreSQL DSN of the pgvector backend; RAG_DATABASE_DSN is used when it is unset.
OPENSEARCH_HOSTS	http://localhost:9200	Comma-separated OpenSearch hosts, with OPENSEARCH_USER and OPENSEARCH_PASSWORD.
OPENSEARCH_USE_SSL	false	Connect to OpenSearch over SSL.
OPENSEARCH_VERIFY_CERTS	false	Verify the OpenSearch server's certificate.
OPENSEARCH_TIMEOUT	30	Seconds an OpenSearch request may take.
OPENSEARCH_ENABLE_NEURAL_SPARSE	false	Add neural sparse vectors to OpenSearch hybrid search (dense, sparse and BM25).

Paths and the process

Variable	Default	Description
LASAGNA_DATA_DIR, RAG_DATA_DIR	~/.lasagna	Where Lasagna keeps its data.
LASAGNA_DB_DIR, RAG_DB_DIR	/dbs	The databases and the local vector stores.
LASAGNA_UPLOAD_DIR, RAG_UPLOAD_DIR	/uploads	Stored uploads.
LASAGNA_SETTINGS_FILE, RAG_SETTINGS_FILE	the packaged settings.lasagna.toml	The settings file to read instead of the packaged one; a path with no file there stops the process with an error.
LASAGNA_LLM_CACHE	on	0, false, no or off turns the LLM response cache off for every run.
RAG_STORE_LOCK_TIMEOUT	300	Seconds a process waits for a lock another process holds in a PostgreSQL metadata store (a document's during a reindex, a session's during a profile switch) before the operation fails and says so, with nothing changed.
LASAGNA_STRICT_PLUGINS	unset	Set: a plugin that fails to load raises instead of being reported and skipped.
RAG_LANGFUSE_TRACING	false	Send LLM traces to Langfuse (needs the langfuse package and its keys).
RAG_LLM_CACHE_MAX_SIZE	50	How many LLM clients the process keeps for reuse.
RAG_CRAG_GRADING_CONCURRENCY	4	Parallel relevance-grading calls in corrective RAG (CRAG).