feat: retake summary using NousResearch/Hermes-3-Llama-3.1-8B model (#415)

This feature a new modal endpoint, and a complete new way to build the
summary.

## SummaryBuilder

The summary builder is based on conversational model, where an exchange
between the model and the user is made. This allow more context
inclusion and a better respect of the rules.

It requires an endpoint with OpenAI-like completions endpoint
(/v1/chat/completions)

## vLLM Hermes3

Unlike previous deployment, this one use vLLM, which gives OpenAI-like
completions endpoint out of the box. It could also handle guided JSON
generation, so jsonformer is not needed. But, the model is quite good to
follow JSON schema if asked in the prompt.

## Conversion of long/short into summary builder

The builder is identifying participants, find key subjects, get a
summary for each, then get a quick recap.

The quick recap is used as a short_summary, while the markdown including
the quick recap + key subjects + summaries are used for the
long_summary.

This is why the nextjs component has to be updated, to correctly style
h1 and keep the new line of the markdown.

server/reflector/llm/base.py

@@ -156,6 +156,27 @@ class LLM:
 
         return result
 
+    async def completion(
+        self, messages: list, logger: reflector_logger, **kwargs
+    ) -> dict:
+        """
+        Use /v1/chat/completion Open-AI compatible endpoint from the URL
+        It's up to the user to validate anything or transform the result
+        """
+        logger.info("LLM completions", messages=messages)
+
+        try:
+            with self.m_generate.time():
+                result = await retry(self._completion)(messages=messages, **kwargs)
+            self.m_generate_success.inc()
+        except Exception:
+            logger.exception("Failed to call llm after retrying")
+            self.m_generate_failure.inc()
+            raise
+
+        logger.debug("LLM completion result", result=repr(result))
+        return result
+
     def ensure_casing(self, title: str) -> str:
         """
         LLM takes care of word casing, but in rare cases this
@@ -234,6 +255,11 @@ class LLM:
     ) -> str:
         raise NotImplementedError
 
+    async def _completion(
+        self, messages: list, logger: reflector_logger, **kwargs
+    ) -> dict:
+        raise NotImplementedError
+
     def _parse_json(self, result: str) -> dict:
         result = result.strip()
         # try detecting code block if exist

server/reflector/llm/llm_modal.py

@@ -23,7 +23,11 @@ class ModalLLM(LLM):
         """
         # TODO: Query the specific GPU platform
         # Replace this with a HTTP call
-        return ["lmsys/vicuna-13b-v1.5", "HuggingFaceH4/zephyr-7b-alpha"]
+        return [
+            "lmsys/vicuna-13b-v1.5",
+            "HuggingFaceH4/zephyr-7b-alpha",
+            "NousResearch/Hermes-3-Llama-3.1-8B",
+        ]
 
     async def _generate(
         self, prompt: str, gen_schema: dict | None, gen_cfg: dict | None, **kwargs
@@ -53,6 +57,31 @@ class ModalLLM(LLM):
             text = response.json()["text"]
             return text
 
+    async def _completion(self, messages: list, **kwargs) -> dict:
+        kwargs.setdefault("temperature", 0.3)
+        kwargs.setdefault("max_tokens", 2048)
+        kwargs.setdefault("stream", False)
+        kwargs.setdefault("repetition_penalty", 1)
+        kwargs.setdefault("top_p", 1)
+        kwargs.setdefault("top_k", -1)
+        kwargs.setdefault("min_p", 0.05)
+        data = {"messages": messages, "model": self.model_name, **kwargs}
+
+        if self.model_name == "NousResearch/Hermes-3-Llama-3.1-8B":
+            self.llm_url = settings.HERMES_3_8B_LLM_URL + "/v1/chat/completions"
+
+        async with httpx.AsyncClient() as client:
+            response = await retry(client.post)(
+                self.llm_url,
+                headers=self.headers,
+                json=data,
+                timeout=self.timeout,
+                retry_timeout=60 * 5,
+                follow_redirects=True,
+            )
+            response.raise_for_status()
+            return response.json()
+
     def _set_model_name(self, model_name: str) -> bool:
         """
         Set the model name