cltl-llm-language-generation-from-triples

Part of the Leolani social robot framework. This component turns structured output from Leolani's episodic knowledge graph (eKG) — statements, thoughts, questions/answers and knowledge gaps expressed as subject-predicate-object triples — into natural language replies, using a local LLM through Ollama and LangChain.

Instead of hand-written phrase templates, an instruction prompt is built for each triple/thought and sent to the LLM, which paraphrases it into a short, natural reply in the target language.

How it works

• src/cltl/reply_generation/llm_triple_replier.py — LLMTripleReplier, the entry point. It configures a ChatOllama model and a PromptProcessor, reads a JSON file of eKG responses (see data/), and for each response generates and prints the prompts and the model's replies.
• src/cltl/reply_generation/prompts/response_processor.py — PromptProcessor extracts the relevant text (statement, thought, novelty, conflict, subject/complement knowledge gaps, etc.) from an eKG response and pairs it with the matching instruction to build the LLM prompt.
• src/cltl/reply_generation/prompts/instruct.py — Instruct holds the system prompts (instructions) used to tell the LLM how to paraphrase each kind of input (statement, answer, no-answer, subject gap, object gap, novelty, conflict) in a given language.

Data

data/ contains example eKG responses used as input/test fixtures, and a notebook that walks through the generation process:

• thoughts-responses.json, basic-statements-responses.json, basic-questions-responses.json, basic-mentions-responses.json, basic-experiences-responses.json, carl-responses.json, question_response.json — sample eKG responses (statements, thoughts, questions, mentions, experiences) as produced by the Leolani brain.
• test_response.py — a single hardcoded example response used for testing.
• Llama3-ThoughtsLangChain.ipynb — a notebook demonstrating reply generation with ChatOllama and Llama 3.

Requirements

• Python 3.11+
• Ollama running locally with the desired model pulled (e.g. ollama pull llama3.2)

Install the Python dependencies:

pip install -r requirements.txt

Usage

Run the replier against a sample data file:

python src/cltl/reply_generation/llm_triple_replier.py

This loads data/thoughts-responses.json, generates the prompts for each response and prints the LLM's paraphrased replies to the console.

To use LLMTripleReplier in your own code:

from cltl.reply_generation.llm_triple_replier import LLMTripleReplier

replier = LLMTripleReplier(language="Dutch", model_name="llama3.2")
prompts = replier._processor.get_all_prompt_input_from_response(response)
for prompt in prompts:
    reply = replier._ollama_client.invoke(prompt)
    print(reply.content)

License

MIT — see LICENSE.