champ/agent.py

# app/champ/agent.py

from typing import Literal
from langchain.agents import create_agent
from langchain.agents.middleware import dynamic_prompt, ModelRequest
from langchain_huggingface import ChatHuggingFace, HuggingFaceEndpoint
from langchain_community.vectorstores import FAISS as LCFAISS

from opentelemetry import trace

from .prompts import CHAMP_SYSTEM_PROMPT_V12

tracer = trace.get_tracer(__name__)


def _build_retrieval_query(messages) -> str:
    user_turns = []

    for m in messages:
        # LangChain HumanMessage
        if hasattr(m, "type") and m.type == "human":
            user_turns.append(m.text)

    # Fallback: just use last message
    if not user_turns:
        return messages[-1].text

    return " ".join(user_turns[-2:])


def make_prompt_with_context(
    vector_store: LCFAISS, lang: Literal["en", "fr"], k: int = 4, prompt_template: str | None = None
):
    context_store = {"last_retrieved_docs": []}  # shared mutable container

    @dynamic_prompt
    def prompt_with_context(request: ModelRequest) -> str:
        with tracer.start_as_current_span("retrieving documents"):
            retrieval_query = _build_retrieval_query(request.state["messages"])
            fetch_k = 20
            try:
                retrieved_docs = vector_store.max_marginal_relevance_search(
                    retrieval_query,
                    k=k,
                    fetch_k=fetch_k,
                    lambda_mult=0.5,  # 0.0 = diverse, 1.0 = similar; 0.3–0.7 is typical
                )
            except Exception:
                retrieved_docs = vector_store.similarity_search(retrieval_query, k=k)

        seen = set()
        unique_docs = []
        for doc in retrieved_docs:
            text = (doc.page_content or "").strip()
            if not text or text in seen:
                continue
            seen.add(text)
            unique_docs.append(doc)

        docs_content = "\n\n".join(doc.page_content for doc in unique_docs)
        context_store["last_retrieved_docs"] = [doc.page_content for doc in unique_docs]

        language = "English" if lang == "en" else "French"
        template = CHAMP_SYSTEM_PROMPT_V12 if prompt_template is None else prompt_template
        return template.format(
            last_query=retrieval_query,
            context=docs_content,
            language=language,
        )

    return prompt_with_context, context_store


def build_champ_agent(
    vector_store: LCFAISS,
    lang: Literal["en", "fr"],
    repo_id: str = "openai/gpt-oss-20b",
    prompt_template: str | None = None,
):
    # Reducing the temperature and increasing top_p is not recommended, because
    # the model would start answering in a very unnatural manner.
    hf_llm = HuggingFaceEndpoint(
        repo_id=repo_id,
        task="text-generation",
        max_new_tokens=1024,
        temperature=0.2,
        top_p=0.9,
        # huggingfacehub_api_token=... (optional; see service.py)
    )
    # Unforntunately, LangChain and Ecologits do not work togehter.
    model_chat = ChatHuggingFace(llm=hf_llm)
    prompt_middleware, context_store = make_prompt_with_context(vector_store, lang, prompt_template=prompt_template)
    return create_agent(
        model_chat,
        tools=[],
        middleware=[
            prompt_middleware,
        ],
    ), context_store