The Routing Engine

The router is the brain of l3mcore. It decides which expert should answer each prompt using a multi-vector hybrid scoring system.

Initialization (server startup)

On startup, the router pre-calculates three types of vector representations for each expert:

expert "programmer"
├── vector_description:    embed("Expert in writing, reviewing and debugging source code.")
├── vector_centroid:       normalize(mean([embed("code"), embed("python"), ...]))
└── vectors_keywords[]:    [embed("code"), embed("python"), embed("javascript"), ...]

This happens only once at startup. Real-time comparisons are fast because only the incoming prompt needs to be vectorized.

Routing process (per request)

1. Prompt arrives: "Help me debug this JavaScript error"
2. embed(prompt) → vector_prompt
3. For each expert:
   a. sim_max_keyword = max(cosine(vector_prompt, kw) for kw in keywords)
   b. sim_description = cosine(vector_prompt, vector_description)
   c. sim_mean_keyword = cosine(vector_prompt, vector_centroid)
   d. sim_top3_vote = fraction of top-3 keywords with sim > 0.40
   e. score_raw = 0.40*a + 0.30*b + 0.20*c + 0.10*d
4. scores_normalized = softmax(scores_raw / temperature)
5. winner = argmax(scores_normalized)
6. if scores_normalized[winner] >= confidence_threshold:
      → dispatch to winner
   else:
      → trigger keyword fallback

Why Multi-Vector instead of a single embedding

A single average vector of all keywords loses information:

Problem	Solution in l3mcore
Dispersed keywords cancel each other out	max_keyword: any keyword can "win" alone
General intent not captured	description: complete semantic phrase
A rare keyword dominates the vector	mean_keyword + top3_vote: ask for consensus

Recommended Embedding Models

Model	Size	Languages	Notes
intfloat/multilingual-e5-small	~120 MB	100+	Default. Balanced.
intfloat/multilingual-e5-base	~280 MB	100+	Higher accuracy, more RAM
BAAI/bge-small-en-v1.5	~130 MB	English	Excellent if you only use English
nomic-ai/nomic-embed-text-v1	~550 MB	English	High quality, heavier

Changing the model only requires updating model_path in config.json.

classification type Router

For advanced cases, you can train a fine-tuned BERT classifier:

{
  "router_type": "classification",
  "model_path": "path/to/your-finetuned-bert-model"
}

The classification model directly predicts the expert label. It is faster in inference but less flexible — adding a new expert requires re-training the model.

Non-ML Mode (model_path: "")

{
  "router_type": "embedding",
  "model_path": "",
  "keyword_fallback": true
}

Disables the ML router completely and uses only keyword matching with fuzzy matching. Useful for Raspberry Pi or very limited hardware.