config.json — Router Configuration

The config/config.json file controls the global behavior of the routing engine and the server.

Online Web Configurator

To make creating your configuration files easier, you can use the l3mcore Online Configurator to visually and intuitively generate your experts.json file.

Complete Structure

{
    "router": {
        "mode": "generic",
        "router_type": "embedding",
        "model_path": "intfloat/multilingual-e5-small",
        "categories_file": "config/experts.json",
        "confidence_threshold": 0.4,
        "keyword_fallback": true,
        "softmax_temperature": 0.15,
        "scoring_weights": {
            "max_keyword":  0.40,
            "description":  0.30,
            "mean_keyword": 0.20,
            "top3_vote":    0.10
        }
    }
}

Parameter Reference

Parameter	Type	Default	Description
mode	string	"generic"	"generic" uses experts.json. "model" uses a custom classifier.
router_type	string	"embedding"	"embedding" for vector semantic search (recommended). "classification" for fine-tuned BERT model.
model_path	string	"intfloat/multilingual-e5-small"	HuggingFace repo or local path. Empty "" to disable ML and use only keywords.
categories_file	string	"config/experts.json"	Path to the experts file. Must be inside the project (security).
confidence_threshold	float	0.4	Minimum probability to select an expert. If no one reaches it, fallbacks are used.
keyword_fallback	bool	true	Activates fuzzy matching if ML fails or falls below the threshold.
softmax_temperature	float	0.15	Controls the "decisiveness" of the router. Lower = more decisive. Higher = smoother.
scoring_weights	object	see below	Weights of the hybrid scoring formula.
context_messages	int	3	Number of recent user messages to concatenate for cascade routing when the last message has low confidence.
context_max_chars	int	1600	Maximum characters for the concatenated context text, to respect the token window of the embedding model.

Weight Tuning (scoring_weights)

The four weights must add up to 1.0:

Weight	Default	What it controls
max_keyword	0.40	Maximum similarity with any single keyword. Detects precise terms.
description	0.30	Similarity with the expert's description. Captures the general intent.
mean_keyword	0.20	Mean similarity with all keywords. Measures general domain coverage.
top3_vote	0.10	Fraction of the top 3 keywords above 0.4 threshold. Requires consensus.

Tuning Examples

Prioritize exact matches of technical terms:

"scoring_weights": {
    "max_keyword": 0.55,
    "description": 0.20,
    "mean_keyword": 0.20,
    "top3_vote": 0.05
}

Prioritize general semantic understanding:

"scoring_weights": {
    "max_keyword": 0.25,
    "description": 0.50,
    "mean_keyword": 0.20,
    "top3_vote": 0.05
}

Softmax Temperature

The temperature controls how "confident" the router is when making a choice:

• 0.05 – 0.15 → Very decisive router, a single option dominates
• 0.20 – 0.35 → Balanced (recommended)
• 0.50+ → Differences between experts are smoothed out, may cause inaccurate routing

warning

Lowering the temperature too much (< 0.05) can make the router rigid and make bad choices in edge cases. Raise the confidence_threshold instead if you want more strictness.