\u4e3b\u8981\u5206\u4e3a\u77e5\u8bc6\u52a0\u5de5<\/strong>\u548cRAG<\/strong>\u90e8\u5206\u5173\u952e\u6d41\u7a0b\uff1a<\/strong><\/p>\n 1. \u77e5\u8bc6\u52a0\u5de5<\/strong><\/p>\n \u77e5\u8bc6\u52a0\u8f7d -> \u77e5\u8bc6\u5207\u7247 -> \u4fe1\u606f\u62bd\u53d6 -> \u77e5\u8bc6\u52a0\u5de5(embedding\/graph\/keywords) -> \u77e5\u8bc6\u5b58\u50a8<\/p>\n <\/p>\n <\/p>\n <\/p>\n \u5982\u4f55\u6269\u5c55\uff1a<\/p>\n <\/p>\n <\/p>\n <\/p>\n ChunkManager: \u901a\u8fc7\u52a0\u8f7d\u540e\u7684\u77e5\u8bc6\u6570\u636e\uff0c\u6839\u636e\u7528\u6237\u6307\u5b9a\u7684\u5206\u7247\u7b56\u7565\u548c\u5206\u7247\u53c2\u6570\u8def\u7531\u5230\u5bf9\u5e94\u7684\u5206\u7247\u5904\u7406\u5668\u8fdb\u884c\u5206\u914d\u3002<\/p>\n <\/p>\n \u5982\u4f55\u6269\u5c55\uff1a\u5982\u679c\u4f60\u60f3\u5728\u754c\u9762\u4e0a\u81ea\u5b9a\u4e49\u4e00\u4e2a\u65b0\u7684\u5206\u7247\u7b56\u7565<\/p>\n <\/p>\n <\/p>\n <\/p>\n <\/p>\n \u6574\u4e2a\u77e5\u8bc6\u6301\u4e45\u5316\u7edf\u4e00\u5b9e\u73b0\u4e86 <\/p>\n <\/p>\n <\/p>\n 2.\u77e5\u8bc6\u68c0\u7d22<\/strong><\/p>\n question -> rewrite -> similarity_search -> rerank -> context_candidates<\/p>\n \u63a5\u4e0b\u6765\u662f\u77e5\u8bc6\u68c0\u7d22\uff0c\u76ee\u524d\u793e\u533a\u7684\u68c0\u7d22\u903b\u8f91\u4e3b\u8981\u5206\u4e3a\u8fd9\u51e0\u6b65\uff0c\u5982\u679c\u4f60\u8bbe\u7f6e\u4e86\u67e5\u8be2\u6539\u5199\u53c2\u6570\uff0c\u76ee\u524d\u4f1a\u901a\u8fc7\u5927\u6a21\u578b\u7ed9\u4f60\u8fdb\u884c\u4e00\u8f6e\u95ee\u9898\u6539\u5199\uff0c\u7136\u540e\u4f1a\u6839\u636e\u4f60\u7684\u77e5\u8bc6\u52a0\u5de5\u65b9\u5f0f\u8def\u7531\u5230\u5bf9\u5e94\u7684\u68c0\u7d22\u5668\uff0c\u5982\u679c\u4f60\u662f\u901a\u8fc7\u5411\u91cf\u8fdb\u884c\u52a0\u5de5\u7684\uff0c\u90a3\u5c31\u4f1a\u901a\u8fc7EmbeddingRetriever\u8fdb\u884c\u68c0\u7d22\uff0c\u5982\u679c\u4f60\u6784\u5efa\u65b9\u5f0f\u662f\u901a\u8fc7\u77e5\u8bc6\u56fe\u8c31\u6784\u5efa\u7684\uff0c\u5c31\u4f1a\u6309\u7167\u77e5\u8bc6\u56fe\u8c31\u65b9\u5f0f\u8fdb\u884c\u68c0\u7d22\uff0c\u5982\u679c\u4f60\u8bbe\u7f6e\u4e86rerank\u6a21\u578b\uff0c\u4f1a\u7ed9\u7c97\u7b5b\u540e\u7684\u5019\u9009\u503c\u8fdb\u884c\u7cbe\u7b5b\uff0c\u8ba9\u5019\u9009\u503c\u548c\u7528\u6237\u95ee\u9898\u66f4\u6709\u5173\u8054\u3002<\/p>\n <\/p>\n![](\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/4a23debbd60c739.png\")
<\/p>\n![](\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/2d848bb79485a36.png\")
<\/p>\n\n
# \u77e5\u8bc6\u5de5\u5382\u8fdb\u884c\u5b9e\u4f8b\u5316\r\nKnowledgeFactory -> create() -> load() -> Document\r\n- knowledge\r\n- markdown\r\n- pdf\r\n- docx\r\n- txt\r\n- html\r\n- pptx\r\n- url\r\n- ...<\/pre>\n
from abc import ABC\r\nfrom typing import List, Any\r\n\r\nclass Knowledge(ABC):\r\ndef load(self) -> List[Document]:\r\n\"\"\"Load knowledge from data loader.\"\"\"\r\npass\r\n\r\n@classmethod\r\ndef document_type(cls) -> Any:\r\n\"\"\"Get document type.\"\"\"\r\npass\r\n\r\n@classmethod\r\ndef support_chunk_strategy(cls) -> List[ChunkStrategy]:\r\n\"\"\"Return supported chunk strategy.\"\"\"\r\nreturn [\r\nChunkStrategy.CHUNK_BY_SIZE,\r\nChunkStrategy.CHUNK_BY_PAGE,\r\nChunkStrategy.CHUNK_BY_PARAGRAPH,\r\nChunkStrategy.CHUNK_BY_MARKDOWN_HEADER,\r\nChunkStrategy.CHUNK_BY_SEPARATOR,\r\n]\r\n\r\n@classmethod\r\ndef default_chunk_strategy(cls) -> ChunkStrategy:\r\n\"\"\"\r\nReturn default chunk strategy.\r\n\r\nReturns:\r\nChunkStrategy: default chunk strategy\r\n\"\"\"\r\nreturn ChunkStrategy.CHUNK_BY_SIZE<\/pre>\n
\n
![](\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/df8dbdb45a3d866.png\")
<\/p>\nclass ChunkManager:\r\n\"\"\"Manager for chunks.\"\"\"\r\n\r\ndef __init__(\r\nself,\r\nknowledge: Knowledge,\r\nchunk_parameter: Optional[ChunkParameters] = None,\r\nextractor: Optional[Extractor] = None,\r\n):\r\n\"\"\"\r\nCreate a new ChunkManager with the given knowledge.\r\n\r\nArgs:\r\nknowledge: (Knowledge) Knowledge datasource.\r\nchunk_parameter: (Optional[ChunkParameters]) Chunk parameter.\r\nextractor: (Optional[Extractor]) Extractor to use for summarization.\r\n\"\"\"\r\nself._knowledge = knowledge\r\nself._extractor = extractor\r\nself._chunk_parameters = chunk_parameter or ChunkParameters()\r\nself._chunk_strategy = (\r\nchunk_parameter.chunk_strategy\r\nif chunk_parameter and chunk_parameter.chunk_strategy\r\nelse self._knowledge.default_chunk_strategy().name\r\n)\r\nself._text_splitter = self._chunk_parameters.text_splitter\r\nself._splitter_type = self._chunk_parameters.splitter_type<\/pre>\n
\n
class ChunkStrategy(Enum):\r\n\"\"\"Chunk Strategy Enum.\"\"\"\r\n\r\nCHUNK_BY_SIZE: _STRATEGY_ENUM_TYPE = (\r\nRecursiveCharacterTextSplitter,\r\n[\r\n{\r\n\"param_name\": \"chunk_size\",\r\n\"param_type\": \"int\",\r\n\"default_value\": 512,\r\n\"description\": \"The size of the data chunks used in processing.\",\r\n},\r\n{\r\n\"param_name\": \"chunk_overlap\",\r\n\"param_type\": \"int\",\r\n\"default_value\": 50,\r\n\"description\": \"The amount of overlap between adjacent data chunks.\",\r\n},\r\n],\r\n\"chunk size\",\r\n\"split document by chunk size\",\r\n)\r\n\r\nCHUNK_BY_PAGE: _STRATEGY_ENUM_TYPE = (\r\nPageTextSplitter,\r\n[],\r\n\"page\",\r\n\"split document by page\",\r\n)\r\n\r\nCHUNK_BY_PARAGRAPH: _STRATEGY_ENUM_TYPE = (\r\nParagraphTextSplitter,\r\n[\r\n{\r\n\"param_name\": \"separator\",\r\n\"param_type\": \"string\",\r\n\"default_value\": \"\\n\",\r\n\"description\": \"paragraph separator\",\r\n}\r\n],\r\n\"paragraph\",\r\n\"split document by paragraph\",\r\n)\r\n\r\nCHUNK_BY_SEPARATOR: _STRATEGY_ENUM_TYPE = (\r\nSeparatorTextSplitter,\r\n[\r\n{\r\n\"param_name\": \"separator\",\r\n\"param_type\": \"string\",\r\n\"default_value\": \"\\n\",\r\n\"description\": \"chunk separator\",\r\n},\r\n{\r\n\"param_name\": \"enable_merge\",\r\n\"param_type\": \"boolean\",\r\n\"default_value\": False,\r\n\"description\": (\r\n\"Whether to merge according to the chunk_size after \"\r\n\"splitting by the separator.\"\r\n),\r\n},\r\n],\r\n\"separator\",\r\n\"split document by separator\",\r\n)\r\n\r\nCHUNK_BY_MARKDOWN_HEADER: _STRATEGY_ENUM_TYPE = (\r\nMarkdownHeaderTextSplitter,\r\n[],\r\n\"markdown header\",\r\n\"split document by markdown header\",\r\n)<\/pre>\n\n
Embeddings<\/code>\u63a5\u53e3<\/li>\n<\/ul>\n@abstractmethod\r\ndef embed_documents(self, texts: List[str]) -> List[List[float]]:\r\n\"\"\"Embed search docs.\"\"\"\r\n\r\n@abstractmethod\r\ndef embed_query(self, text: str) -> List[float]:\r\n\"\"\"Embed query text.\"\"\"\r\n\r\nasync def aembed_documents(self, texts: List[str]) -> List[List[float]]:\r\n\"\"\"Asynchronous Embed search docs.\"\"\"\r\nreturn await asyncio.get_running_loop().run_in_executor(\r\nNone, self.embed_documents, texts\r\n)\r\n\r\nasync def aembed_query(self, text: str) -> List[float]:\r\n\"\"\"Asynchronous Embed query text.\"\"\"\r\nreturn await asyncio.get_running_loop().run_in_executor(\r\nNone, self.embed_query, text\r\n)<\/pre>\n
# EMBEDDING_MODEL=proxy_openai\r\n# proxy_openai_proxy_server_url=https:\/\/api.openai.com\/v1\r\n# proxy_openai_proxy_api_key={your-openai-sk}\r\n# proxy_openai_proxy_backend=text-embedding-ada-002\r\n\r\n## qwen embedding model, See dbgpt\/model\/parameter.py\r\n# EMBEDDING_MODEL=proxy_tongyi\r\n# proxy_tongyi_proxy_backend=text-embedding-v1\r\n# proxy_tongyi_proxy_api_key={your-api-key}\r\n\r\n## qianfan embedding model, See dbgpt\/model\/parameter.py\r\n# EMBEDDING_MODEL=proxy_qianfan\r\n# proxy_qianfan_proxy_backend=bge-large-zh\r\n# proxy_qianfan_proxy_api_key={your-api-key}\r\n# proxy_qianfan_proxy_api_secret={your-secret-key}<\/pre>\n\n
class TripletExtractor(LLMExtractor):\r\n\"\"\"TripletExtractor class.\"\"\"\r\n\r\ndef __init__(self, llm_client: LLMClient, model_name: str):\r\n\"\"\"Initialize the TripletExtractor.\"\"\"\r\nsuper().__init__(llm_client, model_name, TRIPLET_EXTRACT_PT)\r\n\r\nTRIPLET_EXTRACT_PT = (\r\n\"Some text is provided below. Given the text, \"\r\n\"extract up to knowledge triplets as more as possible \"\r\n\"in the form of (subject, predicate, object).\\n\"\r\n\"Avoid stopwords. The subject, predicate, object can not be none.\\n\"\r\n\"---------------------\\n\"\r\n\"Example:\\n\"\r\n\"Text: Alice is Bob's mother.\\n\"\r\n\"Triplets:\\n(Alice, is mother of, Bob)\\n\"\r\n\"Text: Alice has 2 apples.\\n\"\r\n\"Triplets:\\n(Alice, has 2, apple)\\n\"\r\n\"Text: Alice was given 1 apple by Bob.\\n\"\r\n\"Triplets:(Bob, gives 1 apple, Alice)\\n\"\r\n\"Text: Alice was pushed by Bob.\\n\"\r\n\"Triplets:(Bob, pushes, Alice)\\n\"\r\n\"Text: Bob's mother Alice has 2 apples.\\n\"\r\n\"Triplets:\\n(Alice, is mother of, Bob)\\n(Alice, has 2, apple)\\n\"\r\n\"Text: A Big monkey climbed up the tall fruit tree and picked 3 peaches.\\n\"\r\n\"Triplets:\\n(monkey, climbed up, fruit tree)\\n(monkey, picked 3, peach)\\n\"\r\n\"Text: Alice has 2 apples, she gives 1 to Bob.\\n\"\r\n\"Triplets:\\n\"\r\n\"(Alice, has 2, apple)\\n(Alice, gives 1 apple, Bob)\\n\"\r\n\"Text: Philz is a coffee shop founded in Berkeley in 1982.\\n\"\r\n\"Triplets:\\n\"\r\n\"(Philz, is, coffee shop)\\n(Philz, founded in, Berkeley)\\n\"\r\n\"(Philz, founded in, 1982)\\n\"\r\n\"---------------------\\n\"\r\n\"Text: {text}\\n\"\r\n\"Triplets:\\n\"\r\n)<\/pre>\n\n
\n
\n
IndexStoreBase<\/code>\u63a5\u53e3\uff0c\u76ee\u524d\u63d0\u4f9b\u4e86\u5411\u91cf\u6570\u636e\u5e93\u3001\u56fe\u6570\u636e\u5e93\u3001\u5168\u6587\u7d22\u5f15\u4e09\u7c7b\u5b9e\u73b0<\/p>\n![](\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/0e65020b817fc8e.png\")
<\/p>\n\n
# Base class hierarchy\r\n- VectorStoreBase\r\n- ChromaStore\r\n- MilvusStore\r\n- OceanbaseStore\r\n- ElasticsearchStore\r\n- PGVectorStore\r\n\r\n# Base class definition\r\nclass VectorStoreBase(IndexStoreBase, ABC):\r\n\"\"\"\r\nVector store base class.\r\n\"\"\"\r\n\r\n@abstractmethod\r\ndef load_document(self, chunks: List[Chunk]) -> List[str]:\r\n\"\"\"\r\nLoad document in index database.\r\n\"\"\"\r\npass\r\n\r\n@abstractmethod\r\nasync def aload_document(self, chunks: List[Chunk]) -> List[str]:\r\n\"\"\"\r\nLoad document in index database asynchronously.\r\n\"\"\"\r\npass\r\n\r\n@abstractmethod\r\ndef similar_search_with_scores(\r\nself,\r\ntext: str,\r\ntopk: int,\r\nscore_threshold: float,\r\nfilters: Optional[MetadataFilters] = None,\r\n) -> List[Chunk]:\r\n\"\"\"\r\nPerform a similar search with scores in the index database.\r\n\"\"\"\r\npass\r\n\r\ndef similar_search(\r\nself,\r\ntext: str,\r\ntopk: int,\r\nfilters: Optional[MetadataFilters] = None,\r\n) -> List[Chunk]:\r\n\"\"\"\r\nPerform a similar search in the index database.\r\n\"\"\"\r\nreturn self.similar_search_with_scores(text, topk, 1.0, filters)<\/pre>\n
\n
TuGraphStore<\/code>\u4f1a\u6839\u636e\u4e09\u5143\u7ec4\u751f\u6210\u5177\u4f53\u7684Cypher\u8bed\u53e5\u5e76\u6267\u884c\u3002<\/li>\n<\/ul>\n\n
\n
MemoryGraphStore<\/code>\u548cTuGraphStore<\/code>\u7684\u5b9e\u73b0\uff0c\u6211\u4eec\u4e5f\u63d0\u4f9bNeo4j\u63a5\u53e3\u7ed9\u5f00\u53d1\u8005\u8fdb\u884c\u63a5\u5165\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n# GraphStoreBase -> TuGraphStore -> Neo4jStore\r\ndef insert_triplet(self, subj: str, rel: str, obj: str) -> None:\r\n\"\"\"Add triplet.\"\"\"\r\n# Create queries to merge nodes and relationship\r\nsubj_query = f\"MERGE (n1:{self._node_label} {{id:'{subj}'}})\"\r\nobj_query = f\"MERGE (n2:{self._node_label} {{id:'{obj}'}})\"\r\nrel_query = (\r\nf\"MERGE (n1:{self._node_label} {{id:'{subj}'}})\"\r\nf\"-[r:{self._edge_label} {{id:'{rel}'}}]->\"\r\nf\"(n2:{self._node_label} {{id:'{obj}'}})\"\r\n)\r\n# Execute queries\r\nself.conn.run(query=subj_query)\r\nself.conn.run(query=obj_query)\r\nself.conn.run(query=rel_query)<\/pre>\n\n
\n
{\r\n\"analysis\": {\r\n\"analyzer\": {\r\n\"default\": {\r\n\"type\": \"standard\"\r\n}\r\n}\r\n},\r\n\"similarity\": {\r\n\"custom_bm25\": {\r\n\"type\": \"BM25<\/a>\",\r\n\"k1\": self._k1,\r\n\"b\": self._b\r\n}\r\n}\r\n}\r\n\r\nself._es_mappings = {\r\n\"properties\": {\r\n\"content\": {\r\n\"type\": \"text\",\r\n\"similarity\": \"custom_bm25\"\r\n},\r\n\"metadata\": {\r\n\"type\": \"keyword\"\r\n}\r\n}\r\n}\r\n\r\n# FullTextStoreBase\r\n# ElasticDocumentStore\r\n# OpenSearchStore<\/pre>\n![](\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/ae4453d80a3a59d.png\")
<\/p>\n\n
class EmbeddingRetriever(BaseRetriever):\r\n\"\"\"Embedding retriever.\"\"\"\r\n\r\ndef __init__(\r\nself,\r\nindex_store: IndexStoreBase,\r\ntop_k: int = 4,\r\nquery_rewrite: Optional[QueryRewrite] = None,\r\nrerank: Optional[Ranker] = None,\r\nretrieve_strategy: Optional[RetrieverStrategy] = RetrieverStrategy.EMBEDDING,\r\n):\r\npass\r\n\r\nasync def _aretrieve_with_score(\r\nself,\r\nquery: str,\r\nscore_threshold: float,\r\nfilters: Optional[MetadataFilters] = None,\r\n) -> List[Chunk]:\r\n\"\"\"\r\nRetrieve knowledge chunks with score.\r\n\r\nArgs:\r\nquery (str): Query text.\r\nscore_threshold (float): Score threshold.\r\nfilters: Metadata filters.\r\n\r\nReturns:\r\nList[Chunk]: List of chunks with score.\r\n\"\"\"\r\nqueries = [query]\r\n\r\nnew_queries = await self._query_rewrite.rewrite(\r\norigin_query=query,\r\ncontext=context,\r\nnums=1\r\n)\r\nqueries.extend(new_queries)\r\n\r\ncandidates_with_score = [\r\nself._similarity_search_with_score(\r\nquery,\r\nscore_threshold,\r\nfilters,\r\nroot_tracer.get_current_span_id()\r\n)\r\nfor query in queries\r\n]\r\n\r\nnew_candidates_with_score = await self._rerank.arank(\r\nnew_candidates_with_score,\r\nquery\r\n)\r\n\r\nreturn new_candidates_with_score<\/pre>\n
\n
\n
Optional[MetadataFilters]<\/code>, \u5143\u6570\u636e\u4fe1\u606f\u8fc7\u6ee4\u5668\uff0c\u4e3b\u8981\u662f\u53ef\u4ee5\u7528\u6765\u524d\u7f6e\u901a\u8fc7\u5c5e\u6027\u4fe1\u606f\u7b5b\u6389\u4e00\u4e9b\u4e0d\u5339\u914d\u7684\u5019\u9009\u4fe1\u606f\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\nfrom enum import Enum\r\nfrom typing import Union, List\r\nfrom pydantic import BaseModel, Field\r\n\r\nclass FilterCondition(str, Enum):\r\n\"\"\"Vector Store Meta data filter conditions.\"\"\"\r\nAND = \"and\"\r\nOR = \"or\"\r\n\r\nclass MetadataFilter(BaseModel):\r\n\"\"\"Meta data filter.\"\"\"\r\nkey: str = Field(\r\n..., \r\ndescription=\"The key of metadata to filter.\"\r\n)\r\noperator<\/a>: FilterOperator = Field(\r\ndefault=FilterOperator.EQ, \r\ndescription=\"The operator of metadata filter.\"\r\n)\r\nvalue: Union[str, int, float, List[str], List[int], List[float]] = Field(\r\n..., \r\ndescription=\"The value of metadata to filter.\"\r\n)<\/pre>\n
![](\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/6c25e3cf7353d51.png\")
<\/p>\n![](\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/9cd2228638b0a2b.png\")
<\/p>\n![](\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/94f329cc1153add.png\")
<\/p>\n![](\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/c20e02d00657e7a.png\")
<\/p>\n![](\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/d4f45f835350a71.png\")
<\/p>\n![](\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/ec3a68936096d57.png\")
<\/h3>\n![](\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/8e5e069ed7a504d.png\")
<\/p>\n![](\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/f290e69a1653c91.png\")
<\/p>\n![](\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/4d263e07ddd5539.png\")
<\/p>\n![](\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/ac388cab1c28a17.png\")
<\/p>\n![](\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/ebafac6b50de02b.png\")
<\/p>\n![](\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/5e4a24544b5c987.png\")
<\/p>\n![](\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/b2ad9fe206e3dfa.png\")
<\/p>\n![](\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/7a23319db973b5e.png\")
<\/p>\n![](\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/0fa7ddb06d03d04.png\")
<\/p>\n![](\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/76d86d0363bc157.png\")
<\/p>\n![](\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/0a59fb72ec9966d.png\")
<\/p>\n![](\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/5f60eb85e6265c6.png\")
<\/p>\n![](\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/1bc351017296eb9.png\")
<\/p>\n![](\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/90d10849214bd3b.png\")
<\/p>\n