Files
nvi/api/llm.py

140 lines
4.3 KiB
Python

"""LLM client — single `chat()` entry point that dispatches to the active provider.
Provider selection is via `settings.llm_provider`. Supported:
- "groq" — hits Groq via the openai SDK (groq_base_url + groq_api_key).
- "anthropic" — native Anthropic SDK.
- "openai" — openai SDK with the standard base_url (also handles vLLM /
any OpenAI-compatible endpoint).
Every call opens a Langfuse `generation` span automatically, tagged with the
model name and populated with token usage from the provider's response.
Spans nest under whatever observation the caller has open, so the trace
hierarchy shows `text_to_sql > llm.chat` with usage on the inner node.
"""
from __future__ import annotations
from dataclasses import dataclass
from functools import lru_cache
from typing import Callable
from anthropic import Anthropic
from openai import OpenAI
from api import langfuse_client as lf
from api.config import get_settings
# ── Provider clients (cached) ──
@lru_cache(maxsize=1)
def _anthropic() -> Anthropic:
return Anthropic(api_key=get_settings().anthropic_api_key)
@lru_cache(maxsize=1)
def _groq() -> OpenAI:
s = get_settings()
return OpenAI(api_key=s.groq_api_key, base_url=s.groq_base_url)
@lru_cache(maxsize=1)
def _openai() -> OpenAI:
s = get_settings()
return OpenAI(api_key=s.openai_api_key, base_url=s.openai_base_url)
# ── Per-provider chat impls ──
@dataclass
class _Call:
text: str
usage: dict[str, int] # {"input": N, "output": M, "total": N+M}
def _chat_anthropic(system: str, user: str, max_tokens: int) -> _Call:
s = get_settings()
msg = _anthropic().messages.create(
model=s.anthropic_model,
max_tokens=max_tokens,
system=system,
messages=[{"role": "user", "content": user}],
)
text = "".join(b.text for b in msg.content if getattr(b, "type", None) == "text").strip()
return _Call(
text=text,
usage={
"input": int(msg.usage.input_tokens),
"output": int(msg.usage.output_tokens),
"total": int(msg.usage.input_tokens + msg.usage.output_tokens),
},
)
def _chat_openai_compat(client: OpenAI, model: str, system: str, user: str, max_tokens: int) -> _Call:
resp = client.chat.completions.create(
model=model,
max_tokens=max_tokens,
messages=[
{"role": "system", "content": system},
{"role": "user", "content": user},
],
)
text = (resp.choices[0].message.content or "").strip()
u = resp.usage
usage = {
"input": int(u.prompt_tokens) if u else 0,
"output": int(u.completion_tokens) if u else 0,
"total": int(u.total_tokens) if u else 0,
}
return _Call(text=text, usage=usage)
def _chat_groq(system: str, user: str, max_tokens: int) -> _Call:
return _chat_openai_compat(_groq(), get_settings().groq_model, system, user, max_tokens)
def _chat_openai(system: str, user: str, max_tokens: int) -> _Call:
return _chat_openai_compat(_openai(), get_settings().openai_model, system, user, max_tokens)
_PROVIDERS: dict[str, Callable[[str, str, int], _Call]] = {
"groq": _chat_groq,
"anthropic": _chat_anthropic,
"openai": _chat_openai,
}
def _active_model() -> str:
s = get_settings()
return {
"groq": s.groq_model,
"anthropic": s.anthropic_model,
"openai": s.openai_model,
}.get(s.llm_provider.lower(), s.llm_provider)
# ── Public surface ──
def chat(*, system: str, user: str, max_tokens: int = 1024, span_name: str = "llm.chat") -> str:
"""Run a single chat completion. Opens a Langfuse generation span around
the call with the model name + token usage. `span_name` is the label
shown in Langfuse — pass something descriptive (e.g. "text_to_sql.gen")
so traces are scannable."""
provider = get_settings().llm_provider.lower()
impl = _PROVIDERS.get(provider)
if impl is None:
raise ValueError(
f"unknown llm_provider {provider!r}; supported: {sorted(_PROVIDERS)}"
)
with lf.span(
span_name,
as_type="generation",
model=_active_model(),
input={"system": system, "user": user},
) as gen:
call = impl(system, user, max_tokens)
gen.update(output=call.text, usage_details=call.usage)
return call.text