ragflow/agent/tools/retrieval.py

#
#  Copyright 2024 The InfiniFlow Authors. All Rights Reserved.
#
#  Licensed under the Apache License, Version 2.0 (the "License");
#  you may not use this file except in compliance with the License.
#  You may obtain a copy of the License at
#
#      http://www.apache.org/licenses/LICENSE-2.0
#
#  Unless required by applicable law or agreed to in writing, software
#  distributed under the License is distributed on an "AS IS" BASIS,
#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#  See the License for the specific language governing permissions and
#  limitations under the License.
#
import asyncio
from functools import partial
import json
import logging
import os
import re
from abc import ABC
from agent.tools.base import ToolParamBase, ToolBase, ToolMeta
from common.constants import LLMType
from api.db.services.doc_metadata_service import DocMetadataService
from common.metadata_utils import apply_meta_data_filter
from api.db.services.knowledgebase_service import KnowledgebaseService, validate_dataset_embedding_models
from api.db.services.llm_service import LLMBundle
from api.db.services.memory_service import MemoryService
from api.db.joint_services import memory_message_service
from api.db.joint_services.tenant_model_service import get_tenant_default_model_by_type, resolve_model_config
from common import settings
from common.connection_utils import timeout
from rag.app.tag import label_question
from rag.prompts.generator import cross_languages, kb_prompt, memory_prompt


def _shared_embedding_id(records, mismatch_message: str):
    """Return a stored embedding id when records share one resolved model.

    Comparison uses ``validate_dataset_embedding_models`` so a tenant_model
    UUID and a legacy ``model@instance@provider`` composite are compatible
    when they refer to the same model. Empty or unresolved names stay
    isolated. The returned value is an original stored ``embd_id`` (UUID or
    composite) for ``resolve_model_config``, not a normalized base name.
    """
    err = validate_dataset_embedding_models(records)
    if err:
        raise Exception(mismatch_message)
    return next((rec.embd_id for rec in records if rec.embd_id), None)


class RetrievalParam(ToolParamBase):
    """
    Define the Retrieval component parameters.
    """

    def __init__(self):
        self.meta: ToolMeta = {
            "name": "search_my_dateset",
            "description": "This tool can be utilized for relevant content searching in the datasets.",
            "parameters": {
                "query": {
                    "type": "string",
                    "description": "The keywords to search the dataset. The keywords should be the most important words/terms(includes synonyms) from the original request.",
                    "default": "",
                    "required": True,
                }
            },
        }
        super().__init__()
        self.function_name = "search_my_dateset"
        self.description = "This tool can be utilized for relevant content searching in the datasets."
        self.similarity_threshold = 0.2
        self.keywords_similarity_weight = 0.5
        self.top_n = 8
        self.rerank_candidates_count = 64
        self.top_k = 1024
        self.dataset_ids = []
        self.kb_ids = []  # Deprecated: keep for backward compatibility
        self.memory_ids = []
        self.kb_vars = []
        self.rerank_id = ""
        self.empty_response = ""
        self.use_kg = False
        self.cross_languages = []
        self.toc_enhance = False
        self.meta_data_filter = {}
        self.document_ids = ""

    def check(self):
        self.check_decimal_float(self.similarity_threshold, "[Retrieval] Similarity threshold")
        self.check_decimal_float(self.keywords_similarity_weight, "[Retrieval] Keyword similarity weight")
        self.check_positive_number(self.top_n, "[Retrieval] Top N")

    def get_input_form(self) -> dict[str, dict]:
        return {
            "query": {"name": "Query", "type": "line"},
            "document_ids": {"name": "Document IDs", "type": "line"},
        }


class Retrieval(ToolBase, ABC):
    component_name = "Retrieval"

    @property
    def _dataset_ids(self):
        """Get dataset IDs with backward compatibility for kb_ids."""
        return self._param.dataset_ids or getattr(self._param, "kb_ids", None) or []

    @staticmethod
    def _normalize_document_ids(value) -> list[str]:
        if value is None or value == "" or value == []:
            return []
        if isinstance(value, list):
            return [str(v) for v in value if v]
        if isinstance(value, str):
            value = value.strip()
            if not value:
                return []
            return [value]
        return []

    def _resolve_document_ids(self, kwargs_document_ids=None) -> list[str]:
        doc_ids = self._normalize_document_ids(kwargs_document_ids)
        if doc_ids:
            return doc_ids
        raw = getattr(self._param, "document_ids", None)
        if raw is None or raw == "" or raw == []:
            return []
        if isinstance(raw, str) and self._canvas.is_reff(raw):
            return self._normalize_document_ids(self._canvas.get_variable_value(raw))
        return self._normalize_document_ids(raw)

    def _resolve_manual_filter(self, flt: dict) -> dict:
        # Return a new dict instead of mutating `flt` in place. The caller
        # passes filters straight out of self._param.meta_data_filter, so
        # mutating them would make later invocations reuse a stale value.
        pat = re.compile(self.variable_ref_patt)
        content = flt.get("value", "")

        def replace(match):
            value = self._canvas.get_variable_value(match.group(1))
            if value is None:
                return ""
            elif isinstance(value, partial):
                return "".join(value())
            elif isinstance(value, str):
                return value
            return json.dumps(value, ensure_ascii=False)

        resolved = dict(flt)
        resolved["value"] = self._replace_template_matches(pat, content, replace)
        return resolved

    async def _retrieve_kb(self, query_text: str, document_ids=None):
        kb_ids: list[str] = []
        for id in self._dataset_ids:
            if id.find("@") < 0:
                kb_ids.append(id)
                continue
            kb_nm = self._canvas.get_variable_value(id)
            # if kb_nm is a list
            kb_nm_list = kb_nm if isinstance(kb_nm, list) else [kb_nm]
            for nm_or_id in kb_nm_list:
                e, kb = KnowledgebaseService.get_by_name(nm_or_id, self._canvas._tenant_id)
                if not e:
                    e, kb = KnowledgebaseService.get_by_id(nm_or_id)
                    if not e:
                        raise Exception(f"Dataset({nm_or_id}) does not exist.")
                kb_ids.append(kb.id)

        filtered_kb_ids: list[str] = list(set([kb_id for kb_id in kb_ids if kb_id]))

        kbs = KnowledgebaseService.get_by_ids(filtered_kb_ids)
        if not kbs:
            raise Exception("No dataset is selected.")

        embd_id = _shared_embedding_id(kbs, "Knowledge bases use different embedding models.")

        embd_mdl = None
        if embd_id:
            tenant_id = self._canvas.get_tenant_id()
            embd_model_config = resolve_model_config(tenant_id, LLMType.EMBEDDING, embd_id)
            embd_mdl = LLMBundle(tenant_id, embd_model_config)

        rerank_mdl = None
        if self._param.rerank_id:
            rerank_model_config = resolve_model_config(kbs[0].tenant_id, LLMType.RERANK, self._param.rerank_id)
            rerank_mdl = LLMBundle(kbs[0].tenant_id, rerank_model_config)

        vars = self.get_input_elements_from_text(query_text)
        vars = {k: o["value"] for k, o in vars.items()}
        query = self.string_format(query_text, vars)

        doc_ids = self._resolve_document_ids(document_ids)
        if doc_ids:
            logging.info("Retrieval resolved document_ids=%s", doc_ids)
            kb_doc_ids = KnowledgebaseService.list_documents_by_ids(filtered_kb_ids)
            for doc_id in doc_ids:
                if doc_id not in kb_doc_ids:
                    raise Exception(f"The datasets don't own the document {doc_id}")
        if self._param.meta_data_filter != {}:
            # Defer the (potentially expensive) metadata table load — manual
            # filters served by ES push-down never need it. The loader is
            # invoked at most once per request by ``apply_meta_data_filter``.
            def _load_metas() -> dict:
                return DocMetadataService.get_flatted_meta_by_kbs(kb_ids)

            chat_mdl = None
            if self._param.meta_data_filter.get("method") in ["auto", "semi_auto"]:
                tenant_id = self._canvas.get_tenant_id()
                chat_model_config = get_tenant_default_model_by_type(tenant_id, LLMType.CHAT)
                chat_mdl = LLMBundle(tenant_id, chat_model_config)

            doc_ids = await apply_meta_data_filter(
                self._param.meta_data_filter,
                None,
                query,
                chat_mdl,
                doc_ids,
                self._resolve_manual_filter if self._param.meta_data_filter.get("method") == "manual" else None,
                kb_ids=kb_ids,
                metas_loader=_load_metas,
            )

        if self._param.cross_languages:
            query = await cross_languages(kbs[0].tenant_id, None, query, self._param.cross_languages)

        if kbs:
            query = re.sub(r"^user[:：\s]*", "", query, flags=re.IGNORECASE)
            kbinfos = await settings.retriever.retrieval(
                query,
                embd_mdl,
                [kb.tenant_id for kb in kbs],
                filtered_kb_ids,
                1,
                self._param.top_n,
                self._param.similarity_threshold,
                1 - self._param.keywords_similarity_weight,
                knn_top_k=self._param.top_k,
                doc_ids=doc_ids,
                aggs=True,
                rerank_mdl=rerank_mdl,
                rank_feature=label_question(query, kbs),
                rerank_candidates_count=self._param.rerank_candidates_count,
            )
            if self.check_if_canceled("Retrieval processing"):
                return

            if self._param.toc_enhance:
                tenant_id = self._canvas._tenant_id
                chat_model_config = get_tenant_default_model_by_type(tenant_id, LLMType.CHAT)
                chat_mdl = LLMBundle(tenant_id, chat_model_config)
                cks = await settings.retriever.retrieval_by_toc(query, kbinfos["chunks"], [kb.tenant_id for kb in kbs], chat_mdl, self._param.top_n)
                if self.check_if_canceled("Retrieval processing"):
                    return
                if cks:
                    kbinfos["chunks"] = cks
            kbinfos["chunks"] = settings.retriever.retrieval_by_children(kbinfos["chunks"], [kb.tenant_id for kb in kbs])
            if self._param.use_kg:
                tenant_id = self._canvas.get_tenant_id()
                chat_model_config = get_tenant_default_model_by_type(tenant_id, LLMType.CHAT)
                ck = await settings.kg_retriever.retrieval(query, [kb.tenant_id for kb in kbs], kb_ids, embd_mdl, LLMBundle(tenant_id, chat_model_config))
                if self.check_if_canceled("Retrieval processing"):
                    return
                if ck["content_with_weight"]:
                    kbinfos["chunks"].insert(0, ck)
        else:
            kbinfos = {"chunks": [], "doc_aggs": []}

        if self._param.use_kg and kbs:
            chat_model_config = get_tenant_default_model_by_type(kbs[0].tenant_id, LLMType.CHAT)
            ck = await settings.kg_retriever.retrieval(query, [kb.tenant_id for kb in kbs], filtered_kb_ids, embd_mdl, LLMBundle(kbs[0].tenant_id, chat_model_config))
            if self.check_if_canceled("Retrieval processing"):
                return
            if ck["content_with_weight"]:
                ck["content"] = ck["content_with_weight"]
                del ck["content_with_weight"]
                kbinfos["chunks"].insert(0, ck)

        for ck in kbinfos["chunks"]:
            if "vector" in ck:
                del ck["vector"]
            if "content_ltks" in ck:
                del ck["content_ltks"]

        if not kbinfos["chunks"]:
            self.set_output("formalized_content", self._param.empty_response)
            return

        # Format the chunks for JSON output (similar to how other tools do it)
        json_output = kbinfos["chunks"].copy()

        self._canvas.add_reference(kbinfos["chunks"], kbinfos["doc_aggs"])
        form_cnt = "\n".join(kb_prompt(kbinfos, 200000, True))

        # Set both formalized content and JSON output
        self.set_output("formalized_content", form_cnt)
        self.set_output("json", json_output)

        return form_cnt

    async def _retrieve_memory(self, query_text: str):
        memory_ids: list[str] = [memory_id for memory_id in self._param.memory_ids]
        user_id: str = self._param.user_id if hasattr(self._param, "user_id") else None
        memory_list = MemoryService.get_by_ids(memory_ids)
        if not memory_list:
            raise Exception("No memory is selected.")

        _shared_embedding_id(memory_list, "Memory use different embedding models.")

        vars = self.get_input_elements_from_text(query_text)
        vars = {k: o["value"] for k, o in vars.items()}
        query = self.string_format(query_text, vars)
        # query message
        filter_dict: dict = {"memory_id": memory_ids}
        if user_id:
            import re

            # is variable
            if re.match(r"^{.*}$", user_id):
                user_id = self._canvas.get_variable_value(user_id)
            filter_dict["user_id"] = user_id
        message_list = memory_message_service.query_message(
            filter_dict, {"query": query, "similarity_threshold": self._param.similarity_threshold, "keywords_similarity_weight": self._param.keywords_similarity_weight, "top_n": self._param.top_n}
        )
        if not message_list:
            self.set_output("formalized_content", self._param.empty_response)
            return ""
        formated_content = "\n".join(memory_prompt(message_list, 200000))
        # set formalized_content output
        self.set_output("formalized_content", formated_content)

        return formated_content

    @timeout(int(os.environ.get("COMPONENT_EXEC_TIMEOUT", 12)))
    async def _invoke_async(self, **kwargs):
        if self.check_if_canceled("Retrieval processing"):
            return
        if not kwargs.get("query"):
            self.set_output("formalized_content", self._param.empty_response)
            return

        document_ids = self._resolve_document_ids(kwargs.get("document_ids"))
        if hasattr(self._param, "retrieval_from") and self._param.retrieval_from == "dataset":
            return await self._retrieve_kb(kwargs["query"], document_ids)
        elif hasattr(self._param, "retrieval_from") and self._param.retrieval_from == "memory":
            return await self._retrieve_memory(kwargs["query"])
        elif self._dataset_ids:
            return await self._retrieve_kb(kwargs["query"], document_ids)
        elif hasattr(self._param, "memory_ids") and self._param.memory_ids:
            return await self._retrieve_memory(kwargs["query"])
        else:
            self.set_output("formalized_content", self._param.empty_response)
            return

    @timeout(int(os.environ.get("COMPONENT_EXEC_TIMEOUT", 12)))
    def _invoke(self, **kwargs):
        return asyncio.run(self._invoke_async(**kwargs))

    def thoughts(self) -> str:
        return """
Keywords: {}
Looking for the most relevant articles.
        """.format(self.get_input().get("query", "-_-!"))