Knob Registry (Raw)

No description in OpenAPI yet.

No description in OpenAPI yet.

No description in OpenAPI yet.

No description in OpenAPI yet.

No description in OpenAPI yet.

No description in OpenAPI yet.

Default checked user-facing corpus IDs for new conversations.

No description in OpenAPI yet.

No description in OpenAPI yet.

No description in OpenAPI yet.

No description in OpenAPI yet.

CLI command. Receives --prompt, --output, --steps, --width, --height.

No description in OpenAPI yet.

No description in OpenAPI yet.

No description in OpenAPI yet.

No description in OpenAPI yet.

No description in OpenAPI yet.

No description in OpenAPI yet.

No description in OpenAPI yet.

No description in OpenAPI yet.

No description in OpenAPI yet.

No description in OpenAPI yet.

No description in OpenAPI yet.

Provider API endpoint

No description in OpenAPI yet.

Display name

Lower = higher priority when multiple have same model.

No description in OpenAPI yet.

No description in OpenAPI yet.

OpenAI vision detail level.

No description in OpenAPI yet.

No description in OpenAPI yet.

No description in OpenAPI yet.

No description in OpenAPI yet.

Force model for vision. Empty=use chat model if it supports vision.

Protocol for OpenAI cloud_direct calls. 'auto' routes codex-only models to Responses.

No description in OpenAPI yet.

No description in OpenAPI yet.

No description in OpenAPI yet.

No description in OpenAPI yet.

No description in OpenAPI yet.

No description in OpenAPI yet.

Trigger deep when message explicitly references past conversation.

recency_weight for deep (higher when user explicitly asks about the past).

top_k when intensity=deep.

Fallback when classifier is uncertain.

Enable smart gating. False=always query Recall when checked.

Use sparse-only for short questions (< 10 tokens) without explicit recall triggers.

top_k when intensity=light.

Show gate decision (intensity, reason) in status bar.

Show raw RecallSignals in debug footer (dev mode).

Skip Recall for greetings, farewells, acknowledgments.

Messages with ≤ this many tokens are skip candidates (only if they match a skip pattern).

Skip Recall for questions that don't reference past context. 'How does auth work?' doesn't need chat history.

Skip Recall when RAG corpora are checked. Assumes user wants code context, not chat history. Default False — let both contribute.

Default recency weight for Recall (recent messages often more relevant).

top_k for standard Recall queries.

No description in OpenAPI yet.

Chat chunks should be smaller than code chunks.

'turn'=one chunk per message, 'sentence'=split by sentence.

Auto-created at first launch. Users never touch this.

Override embedding model. Empty=use global config.

Enable Recall. ON by default.

Enable Recall graph indexing + retrieval (experimental).

No description in OpenAPI yet.

No description in OpenAPI yet.

pgvector recommended (already running).

No description in OpenAPI yet.

No description in OpenAPI yet.

No description in OpenAPI yet.

State 1: No context. Nothing checked or retrieval returned empty.

State 2: RAG only. Code corpora returned results; Recall did not.

State 4: Both. RAG and Recall both returned results.

No description in OpenAPI yet.

State 3: Recall only. Recall returned results; no RAG corpora active.

No description in OpenAPI yet.

No description in OpenAPI yet.

Temperature when nothing is checked (direct chat = more creative)

Max code snippet length in semantic chunk_summaries

Directories to skip when building chunk_summaries

Keywords that, when present in code, skip the chunk

File patterns/extensions to skip

Max API routes to include per chunk_summary

Max symbols to include per chunk_summary

Max length for purpose field in chunk_summaries

Quick tips shown in chunk_summaries builder UI

Overlap lines for AST chunking

Overlap between chunks

Target chunk size (non-whitespace chars)

Chunking strategy (document + code)

Emit chunk ordinal metadata for neighbor-window retrieval.

Emit parent document id metadata for neighbor-window retrieval.

Target size for greedy chunking

Whether to include fenced code blocks in markdown sections.

Max heading level to split on for markdown chunking.

Maximum tokens per chunk - chunks exceeding this are split recursively

Max file size to index (bytes) - files larger than this are skipped

Minimum chunk size

Token overlap between chunks (token-based strategies)

Include imports in chunks

Max recursion depth for recursive chunking.

Whether to keep separators when splitting (recursive strategy).

Separators for recursive chunking, in priority order.

Target tokens per chunk (token-based strategies)

Port for dev backend (Uvicorn)

Port for dev frontend (Vite)

Timeout for dev stack restart operations (seconds)

Timeout for Docker container actions (start/stop/restart)

Timeout for Docker container list (seconds)

Docker socket URL (e.g., unix:///var/run/docker.sock). Leave empty for auto-detection.

Timeout for Docker infrastructure down command (seconds)

Timeout for Docker infrastructure up command (seconds)

Default number of log lines to tail from containers

Include timestamps in Docker logs (1=yes, 0=no)

Timeout for Docker status check (seconds)

When true, the UI auto-syncs embedding_dim from data/models.json when model changes.

Contextual chunk embedding mode. 'late_chunking_local_only' requires local/HF provider backend.

Prefix added before chunk text prior to embedding (stable document context).

Suffix added after chunk text prior to embedding.

Embedding execution backend. 'deterministic' is offline/test-friendly; 'provider' calls real providers.

Batch size for embedding generation

Enable embedding cache

Embedding dimensions

Max tokens per embedding chunk

OpenAI embedding model

Local SentenceTransformer model

MLX-optimized embedding model (used when embedding_type=mlx)

Max retries for embedding API

Embedding API timeout (seconds)

Embedding provider (dynamic - validated against models.json at runtime)

What to do when text exceeds embedding/token limits.

Max tokens per document segment for local late chunking.

Voyage embedding model

Enable chunk_summary enrichment by default

Max chunk_summaries to generate

Enable chunk enrichment

Max chars for enrichment prompt

Min chars for enrichment

Enrichment timeout (seconds)

Baseline results path

Evaluation dataset path

Multi-query variants for evaluation

K used for ndcg_at_10 metric (default 10).

K used for precision_at_5 metric (default 5).

K used for recall_at_10 metric (default 10).

K used for recall_at_20 metric (default 20).

K used for recall_at_5 metric (default 5).

Weight for graph search results (Neo4j)

Fusion method: 'rrf' (Reciprocal Rank Fusion) or 'weighted' (score-based)

Normalize scores to [0,1] before fusion

RRF smoothing constant (higher = more weight to top ranks)

Weight for sparse BM25/FTS search results

Weight for vector search results (pgvector)

Enrichment backend

Disable code enrichment

Model for code enrichment

Ollama enrichment model

Provider backend for gen_model and channel overrides

Max tokens for generation

Primary generation model

CLI generation model

HTTP transport generation model override

MCP transport generation model override

Ollama generation model

Max retries for generation

Generation temperature

Generation timeout (seconds)

Nucleus sampling threshold

Context window for Ollama

Maximum total time to wait for a local (Ollama) generation request to complete (seconds)

Maximum idle time allowed between streamed chunks from local (Ollama) during generation (seconds)

Ollama API URL

OpenAI API base URL override (for proxies)

Edge weight for AST call relationships (function->callee).

Edge weight for AST containment relationships (module->class/function, class->method).

Edge weight for AST import relationships (module->imported_module).

Edge weight for AST inheritance relationships (class->base).

Build lexical graph (Document/Chunk nodes + NEXT_CHUNK relationships)

Chunk node property that stores the embedding vector

Neo4j vector index name for Chunk embeddings (mode='chunk')

Enable graph building during indexing (Neo4j)

Allowed semantic KG entity types produced by extraction.

Build semantic knowledge graph (concept entities + relations) linked to chunks during indexing

LLM model name for semantic KG extraction when semantic_kg_mode='llm' (empty = use generation.enrich_model)

Timeout (seconds) for semantic KG LLM extraction per chunk

Maximum chunks to process for semantic KG extraction per indexing run (0 = disabled)

Maximum semantic concepts to extract per chunk

Maximum semantic relations to create per chunk (heuristic mode)

Minimum length for semantic concept tokens

Semantic KG extraction mode. 'heuristic' is deterministic and test-friendly; 'llm' uses an LLM to extract entities + relations.

Reasoning effort for semantic KG extraction when using OpenAI Responses-compatible models.

Edge weight for semantic concept relations in heuristic fallback mode.

Edge weight for semantic concept relations in LLM mode.

When true in LLM mode, fail semantic KG extraction for a chunk if LLM extraction fails instead of falling back.

When true, semantic KG extraction preserves/uses typed entities (person, org, location, event, concept).

Store chunk embeddings on Chunk nodes for Neo4j vector search (requires dense embeddings)

Timeout waiting for Neo4j vector index ONLINE (seconds)

Neo4j vector similarity function

Wait for the Neo4j vector index to come ONLINE after (re)creating it

When mode='chunk', expand from seed chunks via Entity graph (IN_CHUNK links) to find related chunks

Blend weight for entity-expansion scores relative to seed chunk scores (mode='chunk')

When mode='chunk', include up to N adjacent chunks (NEXT_CHUNK) around each seed hit

When mode='chunk' and Neo4j uses a shared database, overfetch seed hits before filtering by corpus_id

Enable graph search in tri-brid retrieval

Include community-based expansion in graph search

Maximum graph traversal hops

Graph retrieval mode. 'chunk' uses lexical chunk nodes + Neo4j vector index; 'entity' uses the legacy code-entity graph.

Number of results to retrieve from graph search

Community detection algorithm

Entity types to extract and store in graph

Number of results from graph traversal

Include community detection in graph analysis

Maximum traversal hops for graph search

Automatically create per-corpus Neo4j databases when missing (Enterprise).

Neo4j database name

Database isolation mode: 'shared' uses a single Neo4j database (Community-compatible), 'per_corpus' uses a separate Neo4j database per corpus (Enterprise multi-database).

Prefix for per-corpus Neo4j database names when neo4j_database_mode='per_corpus'.

Neo4j password (defaults to NEO4J_PASSWORD env var when unset)

Neo4j connection URI (bolt:// or neo4j://)

Neo4j username

Neo4j chunk-vector query mode. 'auto' prefers runtime-safe defaults and only uses SEARCH where supported.

Relationship types to extract

Max characters to hydrate

Context hydration mode

Stemmer language

BM25 tokenizer type

Optional local embedding throughput override for index-time estimates (tokens/sec).

Excluded file extensions (comma-separated)

Max file size to index (MB)

Batch size for indexing

Parallel workers for indexing

How to ingest very large text files. 'stream' avoids loading entire files into memory.

When large_file_mode='stream', read text files in bounded char blocks (best-effort).

Include column headers when extracting Parquet text

Max characters per extracted Parquet cell (best-effort)

Max characters to extract from a single Parquet file during indexing (best-effort)

Max rows to extract from a single Parquet file during indexing (best-effort)

Extract only text/string-like columns from Parquet files when possible

PostgreSQL connection string (DSN) for pgvector + FTS storage

Skip dense vector indexing

Auto-generate keywords

Score boost for keyword matches

Max discriminative keywords per repo

Min frequency for keyword

Hours between keyword refresh

Bonus for recently modified files

Bonus for GUI/front-end layers

Bonus for indexing/ingestion layers

Intent-to-layer bonus matrix. Keys are query intents, values are layer->multiplier maps.

Bonus for retrieval/API layers

Penalty for vendor/third-party code (negative values apply a penalty)

Allowed Host header values for MCP HTTP (supports wildcard ':*').

Allowed Origin header values for MCP HTTP (supports wildcard ':*').

Default retrieval mode for MCP search/answer tools when not provided.

Default top_k for MCP search/answer tools when not provided.

Enable DNS rebinding protection for MCP HTTP (recommended).

Enable the embedded MCP Streamable HTTP server.

Prefer JSON responses for MCP Streamable HTTP (recommended).

Mount path for the MCP Streamable HTTP endpoint (e.g. /mcp).

Require `Authorization: Bearer $MCP_API_KEY` for MCP HTTP access.

Run MCP Streamable HTTP in stateless mode (recommended).

Snippet chars for reranking input

Cloud reranker model name when mode=cloud (Cohere: rerank-v3.5)

Cloud reranker provider when mode=cloud (cohere, voyage, jina)

Number of candidates to rerank (cloud mode)

Reranker mode: 'cloud' (Cohere/Voyage/Jina API), 'learning' (MLX Qwen3 LoRA learning reranker), 'none' (disabled). Legacy values 'local'/'hf' normalize to 'learning'.

Reranker API timeout (seconds)

Blend weight for reranker scores

Reranker batch size

Max token length for reranker

Hot-reload on model change

Reload check period (seconds)

Number of candidates to rerank (learning mode)

BM25 length normalization (0=no penalty, 1=full penalty, 0.3-0.5 recommended for code)

BM25 term frequency saturation parameter (higher = more weight to term frequency)

Weight for BM25 in hybrid search

Enable chunk_summary-based retrieval

Minimum confidence threshold

Confidence threshold for avg top-5

Confidence threshold for top-1

Dedup key for final results.

Enable MMR diversification when embeddings are available.

Top-k for evaluation runs

Enable multi-query in eval

Confidence threshold for fallback retrieval strategies

Default top-k for search results

Max characters for result hydration

Result hydration mode

Number of final results to return in LangGraph pipeline

Maximum number of query rewrites for LangGraph pipeline

Max chunks to return per file_path (document-aware result shaping).

Maximum number of query rewrites for multi-query expansion

Minimum score threshold for graph leg results (0 disables).

Minimum score threshold for sparse leg results (0 disables). Note: sparse scores are engine-dependent (FTS vs BM25).

Minimum score threshold for vector leg results (0 disables).

MMR lambda (1=query relevance only, 0=diversity only).

Query variants for multi-query

Include adjacent chunks by ordinal for coherence (requires chunk_ordinal metadata).

Enable synonym expansion

RRF rank smoothing constant (higher = more weight to top ranks)

Top-K for dense vector search

Top-K for sparse BM25 search

Custom path to semantic_synonyms.json (default: data/semantic_synonyms.json)

Enable semantic synonym expansion

Weight for vector search

Bonus score for chunks matched via chunk_summary-based retrieval

Score multiplier when filename exactly matches query terms

Score multiplier when path components match query terms

Comma-separated path prefixes to boost

Vendor code preference

Bypass chat generation cache when images are attached.

Number of prior conversation turns included in chat cache fingerprint.

Enable semantic cache reads/writes (0=off, 1=on).

Maximum cache rows to retain per scope/endpoint.

Skip generation-cache writes when temperature exceeds this value.

Minimum query length before cache is eligible.

Cache mode when enabled.

Minimum cosine similarity for semantic answer cache hits.

Minimum cosine similarity for semantic chat cache hits.

Minimum cosine similarity for semantic search cache hits.

TTL in seconds for answer cache entries.

TTL in seconds for chat cache entries.

TTL in seconds for search cache entries.

BM25 length normalization (0 = no penalty, 1 = full penalty)

BM25 term frequency saturation (higher = more weight to term frequency)

Enable sparse BM25 search in tri-brid retrieval

Sparse retrieval engine. 'postgres_fts' uses built-in FTS; 'pg_search_bm25' uses ParadeDB pg_search.

If sparse retrieval returns empty, run a file_path-based fallback ranking (best-effort).

Max extracted query terms used for file_path fallback.

Enable sparse highlight payloads when supported (UI later).

How to interpret the sparse query string.

Max extracted query terms used for relaxed sparse fallback.

If sparse retrieval returns empty, retry with a relaxed OR-style query (best-effort).

Number of results to retrieve from sparse search

Extract metadata from code chunks during indexing

Analyze eval regressions with skeptical approach - avoid false explanations

Lightweight chunk_summary generation prompt for faster indexing

Main conversational AI system prompt for answering database questions

Generate query variants for better recall in hybrid search

Optimize user query for code search - expand CamelCase, include API nouns

Generate JSON summaries for code chunks during indexing

Prompt for LLM-assisted semantic KG extraction (typed entities + relations)

Judge prompt for synthetic eval row curation and quality filtering

If true, use fast approximate token counting.

HuggingFace tokenizer name (strategy='huggingface').

Lowercase before tokenization.

Absolute hard limit for tokens per chunk (safety ceiling).

Normalize unicode (NFKC) before tokenization for stability.

Tokenization strategy used for chunking/budgeting.

tiktoken encoding name (strategy='tiktoken').

Include resolved alerts

Alert severities to notify

Alert webhook timeout (seconds)

LangChain/LangSmith API endpoint

LangChain project name

Enable LangChain v2 tracing

LangTrace API host

LangTrace project ID

Logging level

Enable metrics collection

Prometheus metrics port

Auto-enable LangSmith tracing

Number of traces to retain

Trace sampling rate (0.0-1.0)

Enable distributed tracing

Tracing backend mode

Query log file path

Learning reranker backend: auto (prefer MLX Qwen3 on Apple Silicon), mlx_qwen3 (force). Legacy values 'transformers'/'hf' normalize to 'auto'.

Base model to fine-tune for MLX Qwen3 learning reranker

Gradient accumulation steps per optimizer update for MLX Qwen3 learning reranker training

LoRA alpha for MLX Qwen3 learning reranker

LoRA dropout for MLX Qwen3 learning reranker

LoRA rank for MLX Qwen3 learning reranker

Module name suffixes to apply LoRA to (MLX Qwen3)

Negative pairs per positive during learning reranker training

Minimum improvement required to auto-promote (primary metric delta)

Promote trained learning artifact to active path only if primary metric improves

Emit trainer telemetry every N optimizer steps (plus first/final)

Unload MLX learning reranker model after idle seconds (0 = never)

Ragweld agent backend (in-process chat model). Currently: mlx_qwen3

Shipped base model for the ragweld agent (MLX).

Gradient accumulation steps per optimizer update for ragweld agent training.

LoRA alpha for ragweld agent MLX fine-tuning.

LoRA dropout for ragweld agent MLX fine-tuning.

LoRA rank for ragweld agent MLX fine-tuning.

Module name suffixes to apply LoRA to (ragweld agent; MLX Qwen3).

Active ragweld agent adapter artifact path (directory containing adapter.npz + adapter_config.json).

Minimum eval_loss improvement required to auto-promote (baseline_loss - new_loss >= epsilon).

Auto-promote trained ragweld agent adapter only if eval_loss improves.

Adapter reload check period (seconds). 0 = check every request.

Emit ragweld agent trainer telemetry every N optimizer steps (plus first/final).

Training dataset path for the ragweld agent (empty = use evaluation.eval_dataset_path).

Unload ragweld agent model after idle seconds (0 = never).

Training batch size

Training epochs for reranker

Learning rate

Warmup steps ratio

Triplet mining mode

Reset triplets file before mining

Active learning reranker artifact path (MLX adapter directory).

Training triplets file path

Min triplets for training

Triplet mining mode

Default model for chat if not specified in request

Max chat history messages

Show citations list on chat answers

Show confidence badge on chat answers

Show dev/debug footer under chat answers

Show routing trace panel by default

Include reasoning/thinking in streamed responses when supported by model

Streaming response timeout in seconds

Enable streaming responses

Max thinking tokens for Anthropic extended thinking

Editor bind mode

Enable editor embedding

Enable embedded editor

Editor Docker image

Embedded editor port

Grafana authentication mode

Grafana base URL

Grafana dashboard slug

Default Grafana dashboard UID

Enable Grafana embedding

Grafana kiosk mode

Grafana organization ID

Grafana refresh interval

Default pane layout preset applied when opening Learning Reranker Studio

Serialized Dockview layout JSON for Learning Reranker Studio pane persistence

Learning Reranker Studio layout engine selection

Preferred logs renderer for Learning Reranker Studio

Show setup summary row above studio dock layout (1=show, 0=collapsed)

Default bottom dock height percentage for Learning Reranker Studio

Use immersive full-height studio mode for Learning Reranker

Default left dock width percentage for Learning Reranker Studio

Default right dock width percentage for Learning Reranker Studio

Enable Learning Reranker Studio V2 layout and controls

Neural Visualizer trajectory coloring mode (absolute loss vs delta loss)

Maximum telemetry points retained for Neural Visualizer

Global motion intensity multiplier for Neural Visualizer effects

Neural Visualizer quality tier

Reduce Neural Visualizer motion for accessibility/performance

Preferred renderer for Neural Visualizer

Render animated vector field accents in Neural Visualizer

Temporal tail length in seconds for visualizer trajectory effects

Target FPS for Neural Visualizer animation loop

Auto-open browser on start

Runtime environment mode (development uses localhost, production uses deployed URLs)

UI theme mode

Enable vector search in tri-brid retrieval

Minimum similarity score threshold (0 = no threshold)

Number of results to retrieve from vector search