ic

apps/ocp/api/agent/main.py

@@ -0,0 +1,130 @@
+import logging
+from pathlib import Path
+from typing import Any
+
+from langchain_core.messages.ai import AIMessageChunk
+from pydantic import BaseModel, Field
+from banortegpt.storage.azure_storage import AzureStorage
+from banortegpt.vector.qdrant import AsyncQdrant
+from langchain_azure_ai.chat_models import AzureAIChatCompletionsModel
+from langchain_azure_ai.embeddings import AzureAIEmbeddingsModel
+
+import api.context as ctx
+from api.config import config
+
+logger = logging.getLogger(__name__)
+
+parent = Path(__file__).parent
+SYSTEM_PROMPT = (parent / "system_prompt.md").read_text()
+
+
+class get_information(BaseModel):
+    """Search a private repository for information."""
+
+    question: str = Field(..., description="The user question")
+
+
+AZURE_AI_URI = "https://eastus2.api.cognitive.microsoft.com"
+
+
+class MayaOCP:
+    system_prompt = SYSTEM_PROMPT
+    generation_config = {
+        "temperature": config.model_temperature,
+    }
+    message_limit = config.message_limit
+    index = config.vector_index
+    limit = config.search_limit
+    bucket = config.storage_bucket
+
+    search = AsyncQdrant.from_config(config)
+    llm = AzureAIChatCompletionsModel(
+        endpoint=f"{AZURE_AI_URI}/openai/deployments/{config.model}",
+        credential=config.openai_api_key,
+    ).bind_tools([get_information])
+    embedder = AzureAIEmbeddingsModel(
+        endpoint=f"{AZURE_AI_URI}/openai/deployments/{config.embedding_model}",
+        credential=config.openai_api_key,
+    )
+    storage = AzureStorage.from_config(config)
+
+    def __init__(self) -> None:
+        self.tool_map = {"get_information": self.get_information}
+
+    def build_response(self, payloads):
+        preface = ["Recuerda citar las referencias en el formato: texto[1]."]
+
+        template = "------ REFERENCIA {index} ----- \n\n{content}"
+
+        filled_templates = [
+            template.format(index=idx, content=payload.get("content", ""))
+            for idx, payload in enumerate(payloads)
+        ]
+
+        return "\n".join(preface + filled_templates)
+
+    async def get_information(self, question: str):
+        logger.info(
+            f"Embedding question: {question} with model {self.embedder.model_name}"
+        )
+        embedding = await self.embedder.aembed_query(question)
+
+        results = await self.search.semantic_search(
+            embedding=embedding, collection=self.index, limit=self.limit
+        )
+
+        tool_response = self.build_response(results)
+        return tool_response, results
+
+    async def get_shareable_urls(self, metadatas: list):
+        reference_urls = []
+        image_urls = []
+
+        for metadata in metadatas:
+            if file := metadata.get("file"):
+                reference_url = await self.storage.get_file_url(
+                    filename=file,
+                    bucket=self.bucket,
+                    minute_duration=20,
+                    image=False,
+                )
+                reference_urls.append(reference_url)
+            if image_file := metadata.get("image"):
+                image_url = await self.storage.get_file_url(
+                    filename=image_file,
+                    bucket=self.bucket,
+                    minute_duration=20,
+                    image=True,
+                )
+                image_urls.append(image_url)
+
+        return reference_urls, image_urls
+
+    def _generation_config_overwrite(self, overwrites: dict | None) -> dict[str, Any]:
+        generation_config_copy = self.generation_config.copy()
+        if overwrites:
+            for k, v in overwrites.items():
+                generation_config_copy[k] = v
+        return generation_config_copy
+
+    async def stream(self, history, overwrites: dict | None = None):
+        generation_config = self._generation_config_overwrite(overwrites)
+
+        async for chunk in self.llm.astream(input=history, **generation_config):
+            assert isinstance(chunk, AIMessageChunk)
+            if call := chunk.tool_call_chunks:
+                if tool_id := call[0].get("id"):
+                    ctx.tool_id.set(tool_id)
+                if name := call[0].get("name"):
+                    ctx.tool_name.set(name)
+                if args := call[0].get("args"):
+                    ctx.tool_buffer.set(ctx.tool_buffer.get() + args)
+            else:
+                if buffer := chunk.content:
+                    assert isinstance(buffer, str)
+                    ctx.buffer.set(ctx.buffer.get() + buffer)
+                    yield buffer
+
+    async def generate(self, history, overwrites: dict | None = None):
+        generation_config = self._generation_config_overwrite(overwrites)
+        return await self.llm.ainvoke(input=history, **generation_config)