Auto-Retrieval: RAG的智能进化

Give me a summary of the SWE-bench paper

改写查询: summary of the SWE-bench paper过滤参数: {"filters": [{"key": "file_name", "value": "swebench.pdf", "operator": "=="}], "condition": "and"}

from llama_index.indices.managed.llama_cloud import LlamaCloudIndeximport os
index = LlamaCloudIndex(  name="research_papers_page",  project_name="llamacloud_demo",  api_key=os.environ["LLAMA_CLOUD_API_KEY"])
doc_retriever = index.as_retriever(    retrieval_mode="files_via_content",    # retrieval_mode="files_via_metadata",    files_top_k=1)
chunk_retriever = index.as_retriever(    retrieval_mode="chunks",    rerank_top_n=5)

from llama_index.core.prompts import ChatPromptTemplatefrom llama_index.core.vector_stores.types import VectorStoreInfo, VectorStoreQuerySpec, MetadataInfo, MetadataFiltersfrom llama_index.core.retrievers import BaseRetrieverfrom llama_index.core.query_engine import RetrieverQueryEnginefrom llama_index.core import Response
import json
SYS_PROMPT = """Your goal is to structure the user's query to match the request schema provided below.You MUST call the tool in order to generate the query spec.
<< Structured Request Schema >>When responding use a markdown code snippet with a JSON object formatted in the following schema:
{schema_str}
The query string should contain only text that is expected to match the contents of documents. Any conditions in the filter should not be mentioned in the query as well.
Make sure that filters only refer to attributes that exist in the data source.Make sure that filters take into account the descriptions of attributes.Make sure that filters are only used as needed. If there are no filters that should be applied return [] for the filter value.
If the user's query explicitly mentions number of documents to retrieve, set top_k to that number, otherwise do not set top_k.
The schema of the metadata filters in the vector db table is listed below, along with some example metadata dictionaries from relevant rows.The user will send the input query string.
Data Source:```json{info_str}```
Example metadata from relevant chunks:{example_rows}
"""
example_rows_retriever = index.as_retriever(    retrieval_mode="chunks",    rerank_top_n=4)
def get_example_rows_fn(**kwargs):    """Retrieve relevant few-shot examples."""    query_str = kwargs["query_str"]    nodes = example_rows_retriever.retrieve(query_str)    # get the metadata, join them    metadata_list = [n.metadata for n in nodes]
    return "n".join([json.dumps(m) for m in metadata_list])            
# TODO: define function mapping for `example_rows`.chat_prompt_tmpl = ChatPromptTemplate.from_messages(    [        ("system", SYS_PROMPT),        ("user", "{query_str}"),    ],    function_mappings={        "example_rows": get_example_rows_fn    })

## NOTE: this is a dataclass that contains information about the metadatavector_store_info = VectorStoreInfo(    content_info="contains content from various research papers",    metadata_info=[        MetadataInfo(            name="file_name",            type="str",            description="Name of the source paper",        ),    ],)
def auto_retriever_rag(query: str, retriever: BaseRetriever) -> Response:    """Synthesizes an answer to your question by feeding in an entire relevant document as context."""    print(f"> User query string: {query}")    # Use structured predict to infer the metadata filters and query string.    query_spec = llm.structured_predict(        VectorStoreQuerySpec,        chat_prompt_tmpl,        info_str=vector_store_info.json(indent=4),        schema_str=VectorStoreQuerySpec.schema_json(indent=4),        query_str=query    )    # build retriever and query engine    filters = MetadataFilters(filters=query_spec.filters) if len(query_spec.filters) > 0 else None    print(f"> Inferred query string: {query_spec.query}")    if filters:        print(f"> Inferred filters: {filters.json()}")    query_engine = RetrieverQueryEngine.from_args(        retriever,         llm=llm,        response_mode="tree_summarize"    )    # run query    return query_engine.query(query_spec.query)

auto_doc_rag("Give me a summary of the SWE-bench paper") print(str(response))

> User query string: Give me a summary of the SWE-bench paper> Inferred query string: summary of the SWE-bench paper> Inferred filters: {"filters": [{"key": "file_name", "value": "swebench.pdf", "operator": "=="}], "condition": "and"}The construction of SWE-Bench involves a three-stage pipeline:
1. **Repo Selection and Data Scraping**: Pull requests (PRs) are collected from 12 popular open-source Python repositories on GitHub, resulting in approximately 90,000 PRs. These repositories are chosen for their popularity, better maintenance, clear contributor guidelines, and extensive test coverage.
2. **Attribute-Based Filtering**: Candidate tasks are created by selecting merged PRs that resolve a GitHub issue and make changes to the test files of the repository. This indicates that the user likely contributed tests to check whether the issue has been resolved.
3. **Execution-Based Filtering**: For each candidate task, the PR’s test content is applied, and the associated test results are logged before and after the PR’s other content is applied. Tasks are filtered out if they do not have at least one test where its status changes from fail to pass or if they result in installation or runtime errors.
Through these stages, the original 90,000 PRs are filtered down to 2,294 task instances that comprise SWE-Bench.