diff options
Diffstat (limited to 'aisearch/scripts/rag_search.py')
| -rw-r--r-- | aisearch/scripts/rag_search.py | 111 |
1 files changed, 111 insertions, 0 deletions
diff --git a/aisearch/scripts/rag_search.py b/aisearch/scripts/rag_search.py new file mode 100644 index 0000000..20d9863 --- /dev/null +++ b/aisearch/scripts/rag_search.py @@ -0,0 +1,111 @@ +"""This is the main module of the package""" + +import argparse +import asyncio +import os +from functools import partial + +import dspy +import torch +from mem0 import Memory +from mem0.configs.base import MemoryConfig + +from gnais.config import Config +from gnais.search.classification import classify_search, extract_keywords +from gnais.search.corpus import ( + create_ensemble_retriever, + get_chroma_db, + get_docs, + init_chroma_db, +) +from gnais.search.prompts import GN_FACT_EXTRACTION_PROMPT, GN_UPDATE_MEMORY_PROMPT +from gnais.search.rag import rag_search + + +def _digest( + query: str, memory: Memory, user_id: str, corpus_path: str, db_path: str +) -> str: + """Run the full RAG pipeline for a single query and return the answer. + + This is a convenience wrapper around the same logic used by the CLI + script, packaged as a plain synchronous function. + """ + + async def _run() -> str: + + docs = get_docs(corpus_path) + chroma_db = get_chroma_db(embed_model="Qwen/Qwen3-Embedding-0.6B") + decision = classify_search(query).get("decision") + retriever = create_ensemble_retriever( + chroma_db=chroma_db, + docs=docs, + keyword_weight=0.7 if decision == "keyword" else 0.5, + ) + + parts = [] + async for chunk in rag_search( + query=( + extract_keywords(query).get("keywords") + if decision == "keyword" + else query + ), + retriever=retriever, + memory=memory, + user_id=user_id, + ): + parts.append(str(chunk)) + return "".join(parts) + + return asyncio.run(_run()) + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument("query", help="Search query") + parser.add_argument("--user-id", default="test-user", help="Mem0 user identity") + args = parser.parse_args() + + torch.manual_seed(Config.SEED) + if torch.cuda.is_available(): + torch.cuda.manual_seed_all(Config.SEED) + + dspy.configure(lm=Config.DEFAULT_LLM) + + os.environ[f"{Config.MEMORY_MODEL.split('/')[0].upper()}_API_KEY"] = Config.API_KEY + memory_config = MemoryConfig( + custom_fact_extraction_prompt=GN_FACT_EXTRACTION_PROMPT, + custom_update_extraction_prompt=GN_UPDATE_MEMORY_PROMPT, + llm={ + "provider": "litellm", + "config": { + "model": Config.MEMORY_MODEL, + "temperature": 0.1, + "max_tokens": 2_000, + "api_key": Config.API_KEY, + }, + }, + embedder={ + "provider": "huggingface", + "config": { + "model": "Qwen/Qwen3-Embedding-0.6B", + "embedding_dims": 1024, + "model_kwargs": { + "trust_remote_code": True, + "device": "cuda" if torch.cuda.is_available() else "cpu", + }, + }, + }, + vector_store={ + "provider": "chroma", + "config": { + "collection_name": "mem0", + "host": "localhost", + "port": 8001, + }, + }, + ) + memory = Memory(config=memory_config) + digest = partial( + _digest, memory=memory, corpus_path=Config.CORPUS_PATH, db_path=Config.DB_PATH + ) + print(digest(query=args.query, user_id=args.user_id)) |
