Configuration Parameters

Complete reference of all available parameters in config.csv.

Server Parameters

Web Server

MCP Server

LLM Parameters

Core LLM Settings

LLM Cache

Embedded LLM Server

Hardware-Specific LLM Tuning

For RTX 3090 (24GB VRAM)

You can run impressive models with proper configuration:

• DeepSeek-R3-Distill-Qwen-7B: Set llm-server-gpu-layers to 35-40
• Qwen2.5-32B-Instruct (Q4_K_M): Fits with llm-server-gpu-layers to 40-45
• DeepSeek-V3 (with MoE): Set llm-server-n-moe to 2-4 to run even 120B models! MoE only loads active experts
• Optimization: Use llm-server-ctx-size of 8192 for longer contexts

For RTX 4070/4070Ti (12-16GB VRAM)

Mid-range cards work great with quantized models:

• Qwen2.5-14B (Q4_K_M): Set llm-server-gpu-layers to 25-30
• DeepSeek-R3-Distill-Llama-8B: Fully fits with layers at 32
• Tips: Keep llm-server-ctx-size at 4096 to save VRAM

For CPU-Only (No GPU)

Modern CPUs can still run capable models:

• DeepSeek-R3-Distill-Qwen-1.5B: Fast on CPU, great for testing
• Phi-3-mini (3.8B): Excellent CPU performance
• Settings: Set llm-server-mlock to true to prevent swapping
• Parallel: Increase llm-server-parallel to CPU cores -2

Recommended Models (GGUF Format)

• Best Overall: DeepSeek-R3-Distill series (1.5B to 70B)
• Best Small: Qwen2.5-3B-Instruct-Q5_K_M
• Best Medium: DeepSeek-R3-Distill-Qwen-14B-Q4_K_M
• Best Large: DeepSeek-V3, Qwen2.5-32B, or GPT2-120B-GGUF (with MoE enabled)

Pro Tip: The llm-server-n-moe parameter is magic for large models - it enables Mixture of Experts, letting you run 120B+ models on consumer hardware by only loading the experts needed for each token!

Local vs Cloud: A Practical Note

General Bots excels at local deployment - you own your hardware, your data stays private, and there are no recurring costs. However, if you need cloud inference:

Groq is the speed champion - They use custom LPU (Language Processing Unit) chips instead of GPUs, delivering 10x faster inference than traditional cloud providers. Their hardware is purpose-built for transformers, avoiding the general-purpose overhead of NVIDIA GPUs.

This isn’t about market competition - it’s about architecture. NVIDIA GPUs are designed for many tasks, while Groq’s chips do one thing incredibly well: transformer inference. If speed matters and you’re using cloud, Groq is currently the fastest option available.

For local deployment, stick with General Bots and the configurations above. For cloud bursts or when you need extreme speed, consider Groq’s API with these settings:

llm-url,https://api.groq.com/openai/v1
llm-key,your-groq-api-key
llm-model,mixtral-8x7b-32768

Embedding Parameters

Email Parameters

Email Read Tracking

When email-read-pixel is enabled, a 1x1 transparent tracking pixel is automatically injected into HTML emails sent via the API. This allows you to:

• Track when emails are opened
• See how many times an email was opened
• Get the approximate location (IP) and device (user agent) of the reader

API Endpoints for tracking:

Example configuration:

email-read-pixel,true
server-url,https://yourdomain.com

Note: The server-url parameter is used to generate the tracking pixel URL. Make sure it’s accessible from the recipient’s email client.

Privacy considerations: Email tracking should be used responsibly. Consider disclosing tracking in your email footer for transparency.

Theme Parameters

Custom Database Parameters

These parameters configure external database connections for use with BASIC keywords like MariaDB/MySQL connections.

Website Crawling Parameters

Image Generator Parameters

Video Generator Parameters

BotModels Service Parameters

Generator Parameters

Teams Channel Parameters

SMS Parameters

Example: MariaDB Connection

custom-server,db.example.com
custom-port,3306
custom-database,myapp
custom-username,botuser
custom-password,secretpass

Multi-Agent Parameters

Agent-to-Agent (A2A) Communication

Bot Reflection

Reflection Types

Available values for bot-reflection-types:

• ConversationQuality - Analyze conversation quality and user satisfaction
• ResponseAccuracy - Analyze response accuracy and relevance
• ToolUsage - Analyze tool usage effectiveness
• KnowledgeRetrieval - Analyze knowledge retrieval performance
• Performance - Analyze overall bot performance

Example:

bot-reflection-enabled,true
bot-reflection-interval,10
bot-reflection-types,ConversationQuality;ResponseAccuracy;ToolUsage
bot-improvement-auto-apply,false
bot-improvement-threshold,7.0

Memory Parameters

User Memory (Cross-Bot)

Episodic Memory (Context Compaction)

Episodic memory automatically manages conversation context to stay within LLM token limits. When conversation exchanges exceed episodic-memory-threshold, older messages are summarized and only the last episodic-memory-history exchanges are kept in full. See Chapter 03 - Episodic Memory for details.

Model Routing Parameters

These parameters configure multi-model routing for different task types. Requires multiple llama.cpp server instances.

Multi-Model Example

llm-models,default;fast;quality;code
llm-url,http://localhost:8081
model-routing-strategy,auto
model-default,fast
model-fallback-enabled,true
model-fallback-order,quality,fast

Hybrid RAG Search Parameters

General Bots uses hybrid search combining dense (embedding) and sparse (BM25 keyword) search for optimal retrieval. The BM25 implementation is powered by Tantivy, a full-text search engine library similar to Apache Lucene.

BM25 Sparse Search (Tantivy)

BM25 is a keyword-based ranking algorithm that excels at finding exact term matches. It’s powered by Tantivy when the vectordb feature is enabled.

Switching Search Modes

Hybrid Search (Default - Best for most use cases)

bm25-enabled,true
rag-dense-weight,0.7
rag-sparse-weight,0.3

Uses both semantic understanding AND keyword matching. Best for general queries.

Dense Only (Semantic Search)

bm25-enabled,false
rag-dense-weight,1.0
rag-sparse-weight,0.0

Uses only embedding-based search. Faster, good for conceptual/semantic queries where exact words don’t matter.

Sparse Only (Keyword Search)

bm25-enabled,true
rag-dense-weight,0.0
rag-sparse-weight,1.0

Uses only BM25 keyword matching. Good for exact term searches, technical documentation, or when embeddings aren’t available.

BM25 Parameter Tuning

The k1 and b parameters control BM25 behavior:

• bm25-k1 (Term Saturation): Controls how much additional term occurrences contribute to the score

  • Lower values (0.5-1.0): Diminishing returns for repeated terms
  • Higher values (1.5-2.0): More weight to documents with many term occurrences
  • Default 1.2 works well for most content

• bm25-b (Length Normalization): Controls document length penalty

  • 0.0: No length penalty (long documents scored equally)
  • 1.0: Full length normalization (strongly penalizes long documents)
  • Default 0.75 balances length fairness

Tuning for specific content:

# For short documents (tweets, titles)
bm25-b,0.3

# For long documents (articles, manuals)
bm25-b,0.9

# For code search (exact matches important)
bm25-k1,1.5
bm25-stemming,false

Code Sandbox Parameters

Example: Python Sandbox

sandbox-enabled,true
sandbox-runtime,lxc
sandbox-timeout,60
sandbox-memory-mb,512
sandbox-cpu-percent,75
sandbox-network,false
sandbox-python-packages,numpy,pandas,requests,matplotlib
sandbox-allowed-paths,/data,/tmp,/uploads

SSE Streaming Parameters

Parameter Types

Boolean

Values: true or false (case-sensitive)

Number

Integer values, must be within valid ranges:

• Ports: 1-65535
• Tokens: Positive integers
• Percentages: 0-100

Float

Decimal values:

• Thresholds: 0.0 to 1.0
• Weights: 0.0 to 1.0

Path

File system paths:

• Relative: ../../../../data/model.gguf
• Absolute: /opt/models/model.gguf

URL

Valid URLs:

• HTTP: http://localhost:8081
• HTTPS: https://api.example.com

String

Any text value (no quotes needed in CSV)

Email

Valid email format: user@domain.com

Hex Color

HTML color codes: #RRGGBB format

Semicolon-separated

Multiple values separated by semicolons: value1;value2;value3

Comma-separated

Multiple values separated by commas: value1,value2,value3

Required vs Optional

Always Required

• None - all parameters have defaults or are optional

Required for Features

• LLM: llm-model must be set
• Email: email-from, email-server, email-user
• Embeddings: embedding-model for knowledge base
• Custom DB: custom-database if using external database

Configuration Precedence

1. Built-in defaults (hardcoded)
2. config.csv values (override defaults)
3. Environment variables (if implemented, override config)

Special Values

• none - Explicitly no value (for llm-key)
• Empty string - Unset/use default
• false - Feature disabled
• true - Feature enabled

Performance Tuning

For Local Models

llm-server-ctx-size,8192
llm-server-n-predict,2048
llm-server-parallel,4
llm-cache,true
llm-cache-ttl,7200

For Production

llm-server-cont-batching,true
llm-cache-semantic,true
llm-cache-threshold,0.90
llm-server-parallel,8
sse-max-connections,5000

For Low Memory

llm-server-ctx-size,2048
llm-server-n-predict,512
llm-server-mlock,false
llm-server-no-mmap,false
llm-cache,false
sandbox-memory-mb,128

For Multi-Agent Systems

a2a-enabled,true
a2a-timeout,30
a2a-max-hops,5
a2a-retry-count,3
a2a-persist-messages,true
bot-reflection-enabled,true
bot-reflection-interval,10
user-memory-enabled,true

For Hybrid RAG

rag-hybrid-enabled,true
rag-dense-weight,0.7
rag-sparse-weight,0.3
rag-reranker-enabled,true
rag-max-results,10
rag-min-score,0.3
rag-cache-enabled,true
bm25-enabled,true
bm25-k1,1.2
bm25-b,0.75

For Dense-Only Search (Faster)

bm25-enabled,false
rag-dense-weight,1.0
rag-sparse-weight,0.0
rag-max-results,10

For Code Execution

sandbox-enabled,true
sandbox-runtime,lxc
sandbox-timeout,30
sandbox-memory-mb,512
sandbox-network,false
sandbox-python-packages,numpy,pandas,requests

Validation Rules

1. Paths: Model files must exist
2. URLs: Must be valid format
3. Ports: Must be 1-65535
4. Emails: Must contain @ and domain
5. Colors: Must be valid hex format
6. Booleans: Exactly true or false
7. Weights: Must sum to 1.0 (e.g., rag-dense-weight + rag-sparse-weight)