Refine resume parts and fix bugs in retrival using sql (#66)

api/apps/conversation_app.py

@@ -21,20 +21,21 @@ from api.db.services.dialog_service import DialogService, ConversationService
 from api.db import LLMType
 from api.db.services.knowledgebase_service import KnowledgebaseService
 from api.db.services.llm_service import LLMService, LLMBundle
-from api.settings import access_logger
+from api.settings import access_logger, stat_logger
 from api.utils.api_utils import server_error_response, get_data_error_result, validate_request
 from api.utils import get_uuid
 from api.utils.api_utils import get_json_result
+from rag.app.resume import forbidden_select_fields4resume
 from rag.llm import ChatModel
 from rag.nlp import retrievaler
 from rag.nlp.search import index_name
-from rag.utils import num_tokens_from_string, encoder
+from rag.utils import num_tokens_from_string, encoder, rmSpace
 
 
 @manager.route('/set', methods=['POST'])
 @login_required
 @validate_request("dialog_id")
-def set():
+def set_conversation():
     req = request.json
     conv_id = req.get("conversation_id")
     if conv_id:
@@ -96,9 +97,10 @@ def rm():
     except Exception as e:
         return server_error_response(e)
 
+
 @manager.route('/list', methods=['GET'])
 @login_required
-def list():
+def list_convsersation():
     dialog_id = request.args["dialog_id"]
     try:
         convs = ConversationService.query(dialog_id=dialog_id)
@@ -175,6 +177,7 @@ def chat(dialog, messages, **kwargs):
     field_map = KnowledgebaseService.get_field_map(dialog.kb_ids)
     ## try to use sql if field mapping is good to go
     if field_map:
+        stat_logger.info("Use SQL to retrieval.")
         markdown_tbl, chunks = use_sql(question, field_map, dialog.tenant_id, chat_mdl)
         if markdown_tbl:
             return {"answer": markdown_tbl, "retrieval": {"chunks": chunks}}
@@ -186,7 +189,8 @@ def chat(dialog, messages, **kwargs):
         if p["key"] not in kwargs:
             prompt_config["system"] = prompt_config["system"].replace("{%s}" % p["key"], " ")
 
-    kbinfos = retrievaler.retrieval(question, embd_mdl, dialog.tenant_id, dialog.kb_ids, 1, dialog.top_n, dialog.similarity_threshold,
+    kbinfos = retrievaler.retrieval(question, embd_mdl, dialog.tenant_id, dialog.kb_ids, 1, dialog.top_n,
+                                    dialog.similarity_threshold,
                                     dialog.vector_similarity_weight, top=1024, aggs=False)
     knowledges = [ck["content_with_weight"] for ck in kbinfos["chunks"]]
 
@@ -220,32 +224,42 @@ def use_sql(question,field_map, tenant_id, chat_mdl):
 {}
 
 问题：{}
-请写出SQL。
+请写出SQL，且只要SQL，不要有其他说明及文字。
 """.format(
         index_name(tenant_id),
         "\n".join([f"{k}: {v}" for k, v in field_map.items()]),
         question
     )
-    sql = chat_mdl.chat(sys_prompt, [{"role": "user", "content": user_promt}], {"temperature": 0.1})
+    sql = chat_mdl.chat(sys_prompt, [{"role": "user", "content": user_promt}], {"temperature": 0.06})
-    sql = re.sub(r".*?select ", "select ", sql, flags=re.IGNORECASE)
+    stat_logger.info(f"“{question}” get SQL: {sql}")
+    sql = re.sub(r"[\r\n]+", " ", sql.lower())
+    sql = re.sub(r".*?select ", "select ", sql.lower())
     sql = re.sub(r" +", " ", sql)
-    sql = re.sub(r"[;；].*", "", sql)
+    sql = re.sub(r"([;；]|```).*", "", sql)
-    if sql[:len("select ")].lower() != "select ":
+    if sql[:len("select ")] != "select ":
         return None, None
-    if sql[:len("select *")].lower() != "select *":
+    if sql[:len("select *")] != "select *":
         sql = "select doc_id,docnm_kwd," + sql[6:]
+    else:
+        flds = []
+        for k in field_map.keys():
+            if k in forbidden_select_fields4resume:continue
+            if len(flds) > 11:break
+            flds.append(k)
+        sql = "select doc_id,docnm_kwd," + ",".join(flds) + sql[8:]
 
-    tbl = retrievaler.sql_retrieval(sql)
+    stat_logger.info(f"“{question}” get SQL(refined): {sql}")
-    if not tbl: return None, None
+    tbl = retrievaler.sql_retrieval(sql, format="json")
+    if not tbl or len(tbl["rows"]) == 0: return None, None
 
     docid_idx = set([ii for ii, c in enumerate(tbl["columns"]) if c["name"] == "doc_id"])
     docnm_idx = set([ii for ii, c in enumerate(tbl["columns"]) if c["name"] == "docnm_kwd"])
     clmn_idx = [ii for ii in range(len(tbl["columns"])) if ii not in (docid_idx | docnm_idx)]
 
     # compose markdown table
-    clmns = "|".join([re.sub(r"/.*", "", field_map.get(tbl["columns"][i]["name"], f"C{i}")) for i in clmn_idx]) + "|原文"
+    clmns = "|".join([re.sub(r"(/.*|（[^（）]+）)", "", field_map.get(tbl["columns"][i]["name"], f"C{i}")) for i in clmn_idx]) + "|原文"
     line = "|".join(["------" for _ in range(len(clmn_idx))]) + "|------"
-    rows = ["|".join([str(r[i]) for i in clmn_idx])+"|" for r in tbl["rows"]]
+    rows = ["|".join([rmSpace(str(r[i])) for i in clmn_idx]).replace("None", " ") + "|" for r in tbl["rows"]]
     if not docid_idx or not docnm_idx:
         access_logger.error("SQL missing field: " + sql)
         return "\n".join([clmns, line, "\n".join(rows)]), []

api/apps/dialog_app.py

@@ -27,7 +27,7 @@ from api.utils.api_utils import get_json_result
 
 @manager.route('/set', methods=['POST'])
 @login_required
-def set():
+def set_dialog():
     req = request.json
     dialog_id = req.get("dialog_id")
     name = req.get("name", "New Dialog")

api/apps/document_app.py

@@ -262,17 +262,18 @@ def rename():
         return server_error_response(e)
 
 
-@manager.route('/get', methods=['GET'])
+@manager.route('/get/<doc_id>', methods=['GET'])
-@login_required
+def get(doc_id):
-def get():
-    doc_id = request.args["doc_id"]
     try:
         e, doc = DocumentService.get_by_id(doc_id)
         if not e:
             return get_data_error_result(retmsg="Document not found!")
 
-        blob = MINIO.get(doc.kb_id, doc.location)
+        response = flask.make_response(MINIO.get(doc.kb_id, doc.location))
-        return get_json_result(data={"base64": base64.b64decode(blob)})
+        ext = re.search(r"\.([^.]+)$", doc.name)
+        if ext:
+            response.headers.set('Content-Type', 'application/%s'%ext.group(1))
+        return response
     except Exception as e:
         return server_error_response(e)
 

api/apps/kb_app.py

@@ -38,6 +38,9 @@ def create():
         req["id"] = get_uuid()
         req["tenant_id"] = current_user.id
         req["created_by"] = current_user.id
+        e, t = TenantService.get_by_id(current_user.id)
+        if not e: return get_data_error_result(retmsg="Tenant not found.")
+        req["embd_id"] = t.embd_id
         if not KnowledgebaseService.save(**req): return get_data_error_result()
         return get_json_result(data={"kb_id": req["id"]})
     except Exception as e:

api/apps/llm_app.py

@@ -21,11 +21,12 @@ from api.db.services.llm_service import LLMFactoriesService, TenantLLMService, L
 from api.db.services.user_service import TenantService, UserTenantService
 from api.utils.api_utils import server_error_response, get_data_error_result, validate_request
 from api.utils import get_uuid, get_format_time
-from api.db import StatusEnum, UserTenantRole
+from api.db import StatusEnum, UserTenantRole, LLMType
 from api.db.services.knowledgebase_service import KnowledgebaseService
 from api.db.db_models import Knowledgebase, TenantLLM
 from api.settings import stat_logger, RetCode
 from api.utils.api_utils import get_json_result
+from rag.llm import EmbeddingModel, CvModel, ChatModel
 
 
 @manager.route('/factories', methods=['GET'])
@@ -43,16 +44,37 @@ def factories():
 @validate_request("llm_factory", "api_key")
 def set_api_key():
     req = request.json
+    # test if api key works
+    msg = ""
+    for llm in LLMService.query(fid=req["llm_factory"]):
+        if llm.model_type == LLMType.EMBEDDING.value:
+            mdl = EmbeddingModel[req["llm_factory"]](
+                req["api_key"], llm.llm_name)
+            try:
+                arr, tc = mdl.encode(["Test if the api key is available"])
+                if len(arr[0]) == 0 or tc ==0: raise Exception("Fail")
+            except Exception as e:
+                msg += f"\nFail to access embedding model({llm.llm_name}) using this api key."
+        elif llm.model_type == LLMType.CHAT.value:
+            mdl = ChatModel[req["llm_factory"]](
+                req["api_key"], llm.llm_name)
+            try:
+                m, tc = mdl.chat(None, [{"role": "user", "content": "Hello! How are you doing!"}], {"temperature": 0.9})
+                if not tc: raise Exception(m)
+            except Exception as e:
+                msg += f"\nFail to access model({llm.llm_name}) using this api key." + str(e)
+
+    if msg: return get_data_error_result(retmsg=msg)
+
     llm = {
         "tenant_id": current_user.id,
         "llm_factory": req["llm_factory"],
         "api_key": req["api_key"]
     }
-    # TODO: Test api_key
     for n in ["model_type", "llm_name"]:
         if n in req: llm[n] = req[n]
 
-    TenantLLM.insert(**llm).on_conflict("replace").execute()
+    TenantLLMService.filter_update([TenantLLM.tenant_id==llm["tenant_id"], TenantLLM.llm_factory==llm["llm_factory"]], llm)
     return get_json_result(data=True)
 
 
@@ -69,6 +91,7 @@ def my_llms():
 @manager.route('/list', methods=['GET'])
 @login_required
 def list():
+    model_type = request.args.get("model_type")
     try:
         objs = TenantLLMService.query(tenant_id=current_user.id)
         mdlnms = set([o.to_dict()["llm_name"] for o in objs if o.api_key])
@@ -79,6 +102,7 @@ def list():
 
         res = {}
         for m in llms:
+            if model_type and m["model_type"] != model_type: continue
             if m["fid"] not in res: res[m["fid"]] = []
             res[m["fid"]].append(m)
 

api/apps/user_app.py

@@ -24,7 +24,8 @@ from api.db.services.llm_service import TenantLLMService, LLMService
 from api.utils.api_utils import server_error_response, validate_request
 from api.utils import get_uuid, get_format_time, decrypt, download_img
 from api.db import UserTenantRole, LLMType
-from api.settings import RetCode, GITHUB_OAUTH, CHAT_MDL, EMBEDDING_MDL, ASR_MDL, IMAGE2TEXT_MDL, PARSERS
+from api.settings import RetCode, GITHUB_OAUTH, CHAT_MDL, EMBEDDING_MDL, ASR_MDL, IMAGE2TEXT_MDL, PARSERS, API_KEY, \
+    LLM_FACTORY
 from api.db.services.user_service import UserService, TenantService, UserTenantService
 from api.settings import stat_logger
 from api.utils.api_utils import get_json_result, cors_reponse
@@ -204,8 +205,8 @@ def user_register(user_id, user):
         "role": UserTenantRole.OWNER
     }
     tenant_llm = []
-    for llm in LLMService.query(fid="Infiniflow"):
+    for llm in LLMService.query(fid=LLM_FACTORY):
-        tenant_llm.append({"tenant_id": user_id, "llm_factory": "Infiniflow", "llm_name": llm.llm_name, "model_type":llm.model_type, "api_key": "infiniflow API Key"})
+        tenant_llm.append({"tenant_id": user_id, "llm_factory": LLM_FACTORY, "llm_name": llm.llm_name, "model_type":llm.model_type, "api_key": API_KEY})
 
     if not UserService.save(**user):return
     TenantService.save(**tenant)

api/db/db_models.py

@@ -465,7 +465,8 @@ class Knowledgebase(DataBaseModel):
     tenant_id = CharField(max_length=32, null=False)
     name = CharField(max_length=128, null=False, help_text="KB name", index=True)
     description = TextField(null=True, help_text="KB description")
-    permission = CharField(max_length=16, null=False, help_text="me|team")
+    embd_id = CharField(max_length=128, null=False, help_text="default embedding model ID")
+    permission = CharField(max_length=16, null=False, help_text="me|team", default="me")
     created_by = CharField(max_length=32, null=False)
     doc_num = IntegerField(default=0)
     token_num = IntegerField(default=0)

api/db/init_data.py

@@ -46,11 +46,6 @@ def init_llm_factory():
             "logo": "",
             "tags": "LLM,TEXT EMBEDDING,SPEECH2TEXT,MODERATION",
             "status": "1",
-        },{
-            "name": "Infiniflow",
-            "logo": "",
-            "tags": "LLM,TEXT EMBEDDING,SPEECH2TEXT,MODERATION",
-            "status": "1",
         },{
             "name": "智普AI",
             "logo": "",
@@ -135,59 +130,33 @@ def init_llm_factory():
             "model_type": LLMType.SPEECH2TEXT.value
         },{
             "fid": factory_infos[1]["name"],
-            "llm_name": "qwen_vl_chat_v1",
+            "llm_name": "qwen-vl-max",
-            "tags": "LLM,CHAT,IMAGE2TEXT",
-            "max_tokens": 765,
-            "model_type": LLMType.IMAGE2TEXT.value
-        },
-        # ----------------------- Infiniflow -----------------------
-        {
-            "fid": factory_infos[2]["name"],
-            "llm_name": "gpt-3.5-turbo",
-            "tags": "LLM,CHAT,4K",
-            "max_tokens": 4096,
-            "model_type": LLMType.CHAT.value
-        },{
-            "fid": factory_infos[2]["name"],
-            "llm_name": "text-embedding-ada-002",
-            "tags": "TEXT EMBEDDING,8K",
-            "max_tokens": 8191,
-            "model_type": LLMType.EMBEDDING.value
-        },{
-            "fid": factory_infos[2]["name"],
-            "llm_name": "whisper-1",
-            "tags": "SPEECH2TEXT",
-            "max_tokens": 25*1024*1024,
-            "model_type": LLMType.SPEECH2TEXT.value
-        },{
-            "fid": factory_infos[2]["name"],
-            "llm_name": "gpt-4-vision-preview",
             "tags": "LLM,CHAT,IMAGE2TEXT",
             "max_tokens": 765,
             "model_type": LLMType.IMAGE2TEXT.value
         },
         # ---------------------- ZhipuAI ----------------------
         {
-            "fid": factory_infos[3]["name"],
+            "fid": factory_infos[2]["name"],
             "llm_name": "glm-3-turbo",
             "tags": "LLM,CHAT,",
             "max_tokens": 128 * 1000,
             "model_type": LLMType.CHAT.value
         }, {
-            "fid": factory_infos[3]["name"],
+            "fid": factory_infos[2]["name"],
             "llm_name": "glm-4",
             "tags": "LLM,CHAT,",
             "max_tokens": 128 * 1000,
             "model_type": LLMType.CHAT.value
         }, {
-            "fid": factory_infos[3]["name"],
+            "fid": factory_infos[2]["name"],
             "llm_name": "glm-4v",
             "tags": "LLM,CHAT,IMAGE2TEXT",
             "max_tokens": 2000,
             "model_type": LLMType.IMAGE2TEXT.value
         },
         {
-            "fid": factory_infos[3]["name"],
+            "fid": factory_infos[2]["name"],
             "llm_name": "embedding-2",
             "tags": "TEXT EMBEDDING",
             "max_tokens": 512,

api/db/services/knowledgebase_service.py

@@ -77,9 +77,12 @@ class KnowledgebaseService(CommonService):
                 if isinstance(v, dict):
                     assert isinstance(old[k], dict)
                     dfs_update(old[k], v)
+                if isinstance(v, list):
+                    assert isinstance(old[k], list)
+                    old[k] = list(set(old[k]+v))
                 else: old[k] = v
         dfs_update(m.parser_config, config)
-        cls.update_by_id(id, m.parser_config)
+        cls.update_by_id(id, {"parser_config": m.parser_config})
 
 
     @classmethod
@@ -88,6 +91,6 @@ class KnowledgebaseService(CommonService):
         conf = {}
         for k in cls.get_by_ids(ids):
             if k.parser_config and "field_map" in k.parser_config:
-                conf.update(k.parser_config)
+                conf.update(k.parser_config["field_map"])
         return conf
 

api/settings.py

@@ -43,12 +43,14 @@ REQUEST_MAX_WAIT_SEC = 300
 
 USE_REGISTRY = get_base_config("use_registry")
 
-LLM = get_base_config("llm", {})
+LLM = get_base_config("user_default_llm", {})
-CHAT_MDL = LLM.get("chat_model", "gpt-3.5-turbo")
+LLM_FACTORY=LLM.get("factory", "通义千问")
-EMBEDDING_MDL = LLM.get("embedding_model", "text-embedding-ada-002")
+CHAT_MDL = LLM.get("chat_model", "qwen-plus")
-ASR_MDL = LLM.get("asr_model", "whisper-1")
+EMBEDDING_MDL = LLM.get("embedding_model", "text-embedding-v2")
+ASR_MDL = LLM.get("asr_model", "paraformer-realtime-8k-v1")
+IMAGE2TEXT_MDL = LLM.get("image2text_model", "qwen-vl-max")
+API_KEY = LLM.get("api_key", "infiniflow API Key")
 PARSERS = LLM.get("parsers", "general:General,qa:Q&A,resume:Resume,naive:Naive,table:Table,laws:Laws,manual:Manual,book:Book,paper:Paper,presentation:Presentation,picture:Picture")
-IMAGE2TEXT_MDL = LLM.get("image2text_model", "gpt-4-vision-preview")
 
 # distribution
 DEPENDENT_DISTRIBUTION = get_base_config("dependent_distribution", False)

api/utils/file_utils.py

@@ -164,10 +164,10 @@ def thumbnail(filename, blob):
         buffered = BytesIO()
         Image.frombytes("RGB", [pix.width, pix.height],
                         pix.samples).save(buffered, format="png")
-        return "data:image/png;base64," + base64.b64encode(buffered.getvalue())
+        return "data:image/png;base64," + base64.b64encode(buffered.getvalue()).decode("utf-8")
 
     if re.match(r".*\.(jpg|jpeg|png|tif|gif|icon|ico|webp)$", filename):
-        return ("data:image/%s;base64,"%filename.split(".")[-1]) + base64.b64encode(Image.open(BytesIO(blob)).thumbnail((30, 30)).tobytes())
+        return ("data:image/%s;base64,"%filename.split(".")[-1]) + base64.b64encode(Image.open(BytesIO(blob)).thumbnail((30, 30)).tobytes()).decode("utf-8")
 
     if re.match(r".*\.(ppt|pptx)$", filename):
         import aspose.slides as slides
@@ -176,7 +176,7 @@ def thumbnail(filename, blob):
             with slides.Presentation(BytesIO(blob)) as presentation:
                 buffered = BytesIO()
                 presentation.slides[0].get_thumbnail(0.03, 0.03).save(buffered, drawing.imaging.ImageFormat.png)
-                return "data:image/png;base64," + base64.b64encode(buffered.getvalue())
+                return "data:image/png;base64," + base64.b64encode(buffered.getvalue()).decode("utf-8")
         except Exception as e:
             pass
 

conf/mapping.json

@@ -118,11 +118,45 @@
         },
         {
             "dense_vector": {
-              "match": "*_vec",
+              "match": "*_512_vec",
               "mapping": {
                 "type": "dense_vector",
                 "index": true,
-                "similarity": "cosine"
+                "similarity": "cosine",
+                "dims": 512
+              }
+            }
+        },
+        {
+            "dense_vector": {
+              "match": "*_768_vec",
+              "mapping": {
+                "type": "dense_vector",
+                "index": true,
+                "similarity": "cosine",
+                "dims": 768
+              }
+            }
+        },
+        {
+            "dense_vector": {
+              "match": "*_1024_vec",
+              "mapping": {
+                "type": "dense_vector",
+                "index": true,
+                "similarity": "cosine",
+                "dims": 1024
+              }
+            }
+        },
+        {
+            "dense_vector": {
+              "match": "*_1536_vec",
+              "mapping": {
+                "type": "dense_vector",
+                "index": true,
+                "similarity": "cosine",
+                "dims": 1536
               }
             }
         },

conf/service_conf.yaml

@@ -11,7 +11,7 @@ permission:
   dataset: false
 ragflow:
   # you must set real ip address, 127.0.0.1 and 0.0.0.0 is not supported
-  host: 127.0.0.1
+  host: 0.0.0.0
   http_port: 9380
 database:
   name: 'rag_flow'
@@ -21,6 +21,19 @@ database:
   port: 5455
   max_connections: 100
   stale_timeout: 30
+minio:
+  user: 'rag_flow'
+  passwd: 'infini_rag_flow'
+  host: '123.60.95.134:9000'
+es:
+  hosts: 'http://123.60.95.134:9200'
+user_default_llm:
+  factory: '通义千问'
+  chat_model: 'qwen-plus'
+  embedding_model: 'text-embedding-v2'
+  asr_model: 'paraformer-realtime-8k-v1'
+  image2text_model: 'qwen-vl-max'
+  api_key: 'sk-xxxxxxxxxxxxx'
 oauth:
   github:
     client_id: 302129228f0d96055bee

rag/app/book.py

@@ -39,6 +39,11 @@ class Pdf(HuParser):
 
 
 def chunk(filename, binary=None, from_page=0, to_page=100000, callback=None, **kwargs):
+    """
+        Supported file formats are docx, pdf, txt.
+        Since a book is long and not all the parts are useful, if it's a PDF,
+        please setup the page ranges for every book in order eliminate negative effects and save elapsed computing time.
+    """
     doc = {
         "docnm_kwd": filename,
         "title_tks": huqie.qie(re.sub(r"\.[a-zA-Z]+$", "", filename))

rag/app/laws.py

@@ -2,7 +2,6 @@ import copy
 import re
 from io import BytesIO
 from docx import Document
-import numpy as np
 from rag.parser import bullets_category, is_english, tokenize, remove_contents_table, hierarchical_merge, \
     make_colon_as_title
 from rag.nlp import huqie
@@ -59,6 +58,9 @@ class Pdf(HuParser):
 
 
 def chunk(filename, binary=None, from_page=0, to_page=100000, callback=None, **kwargs):
+    """
+        Supported file formats are docx, pdf, txt.
+    """
     doc = {
         "docnm_kwd": filename,
         "title_tks": huqie.qie(re.sub(r"\.[a-zA-Z]+$", "", filename))

rag/app/manual.py

@@ -58,8 +58,10 @@ class Pdf(HuParser):
 
 
 def chunk(filename, binary=None, from_page=0, to_page=100000, callback=None, **kwargs):
+    """
+        Only pdf is supported.
+    """
     pdf_parser = None
-    paper = {}
 
     if re.search(r"\.pdf$", filename, re.IGNORECASE):
         pdf_parser = Pdf()

rag/app/naive.py

@@ -6,6 +6,7 @@ from rag.nlp import huqie
 from rag.parser.pdf_parser import HuParser
 from rag.settings import cron_logger
 
+
 class Pdf(HuParser):
     def __call__(self, filename, binary=None, from_page=0,
                  to_page=100000, zoomin=3, callback=None):
@@ -26,6 +27,12 @@ class Pdf(HuParser):
 
 
 def chunk(filename, binary=None, from_page=0, to_page=100000, callback=None, **kwargs):
+    """
+        Supported file formats are docx, pdf, txt.
+        This method apply the naive ways to chunk files.
+        Successive text will be sliced into pieces using 'delimiter'.
+        Next, these successive pieces are merge into chunks whose token number is no more than 'Max token number'.
+    """
     doc = {
         "docnm_kwd": filename,
         "title_tks": huqie.qie(re.sub(r"\.[a-zA-Z]+$", "", filename))
@@ -45,7 +52,8 @@ def chunk(filename, binary=None, from_page=0, to_page=100000, callback=None, **k
     elif re.search(r"\.txt$", filename, re.IGNORECASE):
         callback(0.1, "Start to parse.")
         txt = ""
-        if binary:txt = binary.decode("utf-8")
+        if binary:
+            txt = binary.decode("utf-8")
         else:
             with open(filename, "r") as f:
                 while True:
@@ -55,10 +63,11 @@ def chunk(filename, binary=None, from_page=0, to_page=100000, callback=None, **k
         sections = txt.split("\n")
         sections = [(l, "") for l in sections if l]
         callback(0.8, "Finish parsing.")
-    else: raise NotImplementedError("file type not supported yet(docx, pdf, txt supported)")
+    else:
+        raise NotImplementedError("file type not supported yet(docx, pdf, txt supported)")
 
-    parser_config = kwargs.get("parser_config", {"chunk_token_num": 128, "delimer": "\n。；！？"})
+    parser_config = kwargs.get("parser_config", {"chunk_token_num": 128, "delimiter": "\n!?。；！？"})
-    cks = naive_merge(sections, parser_config["chunk_token_num"], parser_config["delimer"])
+    cks = naive_merge(sections, parser_config["chunk_token_num"], parser_config["delimiter"])
     eng = is_english(cks)
     res = []
     # wrap up to es documents
@@ -75,6 +84,10 @@ def chunk(filename, binary=None, from_page=0, to_page=100000, callback=None, **k
 
 if __name__ == "__main__":
     import sys
+
+
     def dummy(a, b):
         pass
+
+
     chunk(sys.argv[1], from_page=0, to_page=10, callback=dummy)

rag/app/paper.py

@@ -129,6 +129,10 @@ class Pdf(HuParser):
 
 
 def chunk(filename, binary=None, from_page=0, to_page=100000, callback=None, **kwargs):
+    """
+        Only pdf is supported.
+        The abstract of the paper will be sliced as an entire chunk, and will not be sliced partly.
+    """
     pdf_parser = None
     if re.search(r"\.pdf$", filename, re.IGNORECASE):
         pdf_parser = Pdf()

rag/app/presentation.py

@@ -94,6 +94,11 @@ class Pdf(HuParser):
 
 
 def chunk(filename, binary=None, from_page=0, to_page=100000, callback=None, **kwargs):
+    """
+    The supported file formats are pdf, pptx.
+    Every page will be treated as a chunk. And the thumbnail of every page will be stored.
+    PPT file will be parsed by using this method automatically, setting-up for every PPT file is not necessary.
+    """
     doc = {
         "docnm_kwd": filename,
         "title_tks": huqie.qie(re.sub(r"\.[a-zA-Z]+$", "", filename))

rag/app/qa.py

@@ -70,7 +70,17 @@ def beAdoc(d, q, a, eng):
 
 
 def chunk(filename, binary=None, callback=None, **kwargs):
+    """
+        Excel and csv(txt) format files are supported.
+        If the file is in excel format, there should be 2 column question and answer without header.
+        And question column is ahead of answer column.
+        And it's O.K if it has multiple sheets as long as the columns are rightly composed.
 
+        If it's in csv format, it should be UTF-8 encoded. Use TAB as delimiter to separate question and answer.
+
+        All the deformed lines will be ignored.
+        Every pair of Q&A will be treated as a chunk.
+    """
     res = []
     if re.search(r"\.xlsx?$", filename, re.IGNORECASE):
         callback(0.1, "Start to parse.")

rag/app/resume.py

@@ -4,24 +4,34 @@ import os
 import re
 import requests
 from api.db.services.knowledgebase_service import KnowledgebaseService
+from api.settings import stat_logger
 from rag.nlp import huqie
 
 from rag.settings import cron_logger
 from rag.utils import rmSpace
 
+forbidden_select_fields4resume = [
+    "name_pinyin_kwd", "edu_first_fea_kwd", "degree_kwd", "sch_rank_kwd", "edu_fea_kwd"
+]
 
 def chunk(filename, binary=None, callback=None, **kwargs):
+    """
+    The supported file formats are pdf, docx and txt.
+    To maximize the effectiveness, parse the resume correctly,
+    please visit https://github.com/infiniflow/ragflow, and sign in the our demo web-site
+    to get token. It's FREE!
+    Set INFINIFLOW_SERVER and INFINIFLOW_TOKEN in '.env' file or
+    using 'export' to set both environment variables: INFINIFLOW_SERVER and INFINIFLOW_TOKEN in docker container.
+    """
     if not re.search(r"\.(pdf|doc|docx|txt)$", filename, flags=re.IGNORECASE):
         raise NotImplementedError("file type not supported yet(pdf supported)")
 
     url = os.environ.get("INFINIFLOW_SERVER")
-    if not url:
-        raise EnvironmentError(
-            "Please set environment variable: 'INFINIFLOW_SERVER'")
     token = os.environ.get("INFINIFLOW_TOKEN")
-    if not token:
+    if not url or not token:
-        raise EnvironmentError(
+        stat_logger.warning(
-            "Please set environment variable: 'INFINIFLOW_TOKEN'")
+            "INFINIFLOW_SERVER is not specified. To maximize the effectiveness, please visit https://github.com/infiniflow/ragflow, and sign in the our demo web site to get token. It's FREE! Using 'export' to set both environment variables: INFINIFLOW_SERVER and INFINIFLOW_TOKEN.")
+        return []
 
     if not binary:
         with open(filename, "rb") as f:
@@ -44,22 +54,28 @@ def chunk(filename, binary=None, callback=None, **kwargs):
 
     callback(0.2, "Resume parsing is going on...")
     resume = remote_call()
+    if len(resume.keys()) < 7:
+        callback(-1, "Resume is not successfully parsed.")
+        return []
     callback(0.6, "Done parsing. Chunking...")
     print(json.dumps(resume, ensure_ascii=False, indent=2))
 
     field_map = {
         "name_kwd": "姓名/名字",
+        "name_pinyin_kwd": "姓名拼音/名字拼音",
         "gender_kwd": "性别（男，女）",
         "age_int": "年龄/岁/年纪",
         "phone_kwd": "电话/手机/微信",
         "email_tks": "email/e-mail/邮箱",
         "position_name_tks": "职位/职能/岗位/职责",
-        "expect_position_name_tks": "期望职位/期望职能/期望岗位",
+        "expect_city_names_tks": "期望城市",
+        "work_exp_flt": "工作年限/工作年份/N年经验/毕业了多少年",
+        "corporation_name_tks": "最近就职(上班)的公司/上一家公司",
 
-        "hightest_degree_kwd": "最高学历（高中，职高，硕士，本科，博士，初中，中技，中专，专科，专升本，MPA，MBA，EMBA）",
-        "first_degree_kwd": "第一学历（高中，职高，硕士，本科，博士，初中，中技，中专，专科，专升本，MPA，MBA，EMBA）",
-        "first_major_tks": "第一学历专业",
         "first_school_name_tks": "第一学历毕业学校",
+        "first_degree_kwd": "第一学历（高中，职高，硕士，本科，博士，初中，中技，中专，专科，专升本，MPA，MBA，EMBA）",
+        "highest_degree_kwd": "最高学历（高中，职高，硕士，本科，博士，初中，中技，中专，专科，专升本，MPA，MBA，EMBA）",
+        "first_major_tks": "第一学历专业",
         "edu_first_fea_kwd": "第一学历标签（211，留学，双一流，985，海外知名，重点大学，中专，专升本，专科，本科，大专）",
 
         "degree_kwd": "过往学历（高中，职高，硕士，本科，博士，初中，中技，中专，专科，专升本，MPA，MBA，EMBA）",
@@ -68,14 +84,14 @@ def chunk(filename, binary=None, callback=None, **kwargs):
         "sch_rank_kwd": "学校标签（顶尖学校，精英学校，优质学校，一般学校）",
         "edu_fea_kwd": "教育标签（211，留学，双一流，985，海外知名，重点大学，中专，专升本，专科，本科，大专）",
 
-        "work_exp_flt": "工作年限/工作年份/N年经验/毕业了多少年",
-        "birth_dt": "生日/出生年份",
         "corp_nm_tks": "就职过的公司/之前的公司/上过班的公司",
-        "corporation_name_tks": "最近就职(上班)的公司/上一家公司",
         "edu_end_int": "毕业年份",
-        "expect_city_names_tks": "期望城市",
+        "industry_name_tks": "所在行业",
-        "industry_name_tks": "所在行业"
+
+        "birth_dt": "生日/出生年份",
+        "expect_position_name_tks": "期望职位/期望职能/期望岗位",
     }
+
     titles = []
     for n in ["name_kwd", "gender_kwd", "position_name_tks", "age_int"]:
         v = resume.get(n, "")
@@ -105,6 +121,10 @@ def chunk(filename, binary=None, callback=None, **kwargs):
     doc["content_ltks"] = huqie.qie(doc["content_with_weight"])
     doc["content_sm_ltks"] = huqie.qieqie(doc["content_ltks"])
     for n, _ in field_map.items():
+        if n not in resume:continue
+        if isinstance(resume[n], list) and (len(resume[n]) == 1 or n not in forbidden_select_fields4resume):
+            resume[n] = resume[n][0]
+        if n.find("_tks")>0: resume[n] = huqie.qieqie(resume[n])
         doc[n] = resume[n]
 
     print(doc)

rag/app/table.py

@@ -100,7 +100,20 @@ def column_data_type(arr):
 
 
 def chunk(filename, binary=None, callback=None, **kwargs):
-    dfs = []
+    """
+        Excel and csv(txt) format files are supported.
+        For csv or txt file, the delimiter between columns is TAB.
+        The first line must be column headers.
+        Column headers must be meaningful terms inorder to make our NLP model understanding.
+        It's good to enumerate some synonyms using slash '/' to separate, and even better to
+        enumerate values using brackets like 'gender/sex(male, female)'.
+        Here are some examples for headers:
+            1. supplier/vendor\tcolor(yellow, red, brown)\tgender/sex(male, female)\tsize(M,L,XL,XXL)
+            2. 姓名/名字\t电话/手机/微信\t最高学历（高中，职高，硕士，本科，博士，初中，中技，中专，专科，专升本，MPA，MBA，EMBA）
+
+        Every row in table will be treated as a chunk.
+    """
+
     if re.search(r"\.xlsx?$", filename, re.IGNORECASE):
         callback(0.1, "Start to parse.")
         excel_parser = Excel()
@@ -155,7 +168,7 @@ def chunk(filename, binary=None, callback=None, **kwargs):
                 del df[n]
         clmns = df.columns.values
         txts = list(copy.deepcopy(clmns))
-        py_clmns = [PY.get_pinyins(n)[0].replace("-", "_") for n in clmns]
+        py_clmns = [PY.get_pinyins(re.sub(r"(/.*|（[^（）]+?）|\([^()]+?\))", "", n), '_')[0] for n in clmns]
         clmn_tys = []
         for j in range(len(clmns)):
             cln, ty = column_data_type(df[clmns[j]])

rag/llm/__init__.py

@@ -21,7 +21,7 @@ from .cv_model import *
 EmbeddingModel = {
     "Infiniflow": HuEmbedding,
     "OpenAI": OpenAIEmbed,
-    "通义千问": QWenEmbed,
+    "通义千问": HuEmbedding, #QWenEmbed,
 }
 
 

rag/llm/chat_model.py

@@ -32,7 +32,7 @@ class GptTurbo(Base):
         self.model_name = model_name
 
     def chat(self, system, history, gen_conf):
-        history.insert(0, {"role": "system", "content": system})
+        if system: history.insert(0, {"role": "system", "content": system})
         res = self.client.chat.completions.create(
             model=self.model_name,
             messages=history,
@@ -49,11 +49,12 @@ class QWenChat(Base):
 
     def chat(self, system, history, gen_conf):
         from http import HTTPStatus
-        history.insert(0, {"role": "system", "content": system})
+        if system: history.insert(0, {"role": "system", "content": system})
         response = Generation.call(
             self.model_name,
             messages=history,
-            result_format='message'
+            result_format='message',
+            **gen_conf
         )
         if response.status_code == HTTPStatus.OK:
             return response.output.choices[0]['message']['content'], response.usage.output_tokens
@@ -68,10 +69,11 @@ class ZhipuChat(Base):
 
     def chat(self, system, history, gen_conf):
         from http import HTTPStatus
-        history.insert(0, {"role": "system", "content": system})
+        if system: history.insert(0, {"role": "system", "content": system})
         response = self.client.chat.completions.create(
             self.model_name,
-            messages=history
+            messages=history,
+            **gen_conf
         )
         if response.status_code == HTTPStatus.OK:
             return response.output.choices[0]['message']['content'], response.usage.completion_tokens

rag/llm/embedding_model.py

@@ -100,11 +100,11 @@ class QWenEmbed(Base):
                 input=texts[i:i+batch_size],
                 text_type="document"
             )
-            embds = [[]] * len(resp["output"]["embeddings"])
+            embds = [[] for _ in range(len(resp["output"]["embeddings"]))]
             for e in resp["output"]["embeddings"]:
                 embds[e["text_index"]] = e["embedding"]
             res.extend(embds)
-            token_count += resp["usage"]["input_tokens"]
+            token_count += resp["usage"]["total_tokens"]
         return np.array(res), token_count
 
     def encode_queries(self, text):
@@ -113,7 +113,7 @@ class QWenEmbed(Base):
                 input=text[:2048],
                 text_type="query"
             )
-        return np.array(resp["output"]["embeddings"][0]["embedding"]), resp["usage"]["input_tokens"]
+        return np.array(resp["output"]["embeddings"][0]["embedding"]), resp["usage"]["total_tokens"]
 
 
 from zhipuai import ZhipuAI

rag/nlp/search.py

@@ -92,7 +92,7 @@ class Dealer:
             assert emb_mdl, "No embedding model selected"
             s["knn"] = self._vector(
                 qst, emb_mdl, req.get(
-                    "similarity", 0.4), ps)
+                    "similarity", 0.1), ps)
             s["knn"]["filter"] = bqry.to_dict()
             if "highlight" in s:
                 del s["highlight"]
@@ -106,7 +106,7 @@ class Dealer:
                 bqry.filter.append(Q("terms", kb_id=req["kb_ids"]))
             s["query"] = bqry.to_dict()
             s["knn"]["filter"] = bqry.to_dict()
-            s["knn"]["similarity"] = 0.7
+            s["knn"]["similarity"] = 0.17
             res = self.es.search(s, idxnm=idxnm, timeout="600s", src=src)
 
         kwds = set([])
@@ -171,7 +171,7 @@ class Dealer:
                     continue
                 if not isinstance(v, type("")):
                     m[n] = str(m[n])
-                m[n] = rmSpace(m[n])
+                if n.find("tks")>0: m[n] = rmSpace(m[n])
 
             if m:
                 res[d["id"]] = m
@@ -303,21 +303,22 @@ class Dealer:
 
         return ranks
 
-    def sql_retrieval(self, sql, fetch_size=128):
+    def sql_retrieval(self, sql, fetch_size=128, format="json"):
         sql = re.sub(r"[ ]+", " ", sql)
+        sql = sql.replace("%", "")
+        es_logger.info(f"Get es sql: {sql}")
         replaces = []
-        for r in re.finditer(r" ([a-z_]+_l?tks like |[a-z_]+_l?tks ?= ?)'([^']+)'", sql):
+        for r in re.finditer(r" ([a-z_]+_l?tks)( like | ?= ?)'([^']+)'", sql):
-            fld, v = r.group(1), r.group(2)
+            fld, v = r.group(1), r.group(3)
-            fld = re.sub(r" ?(like|=)$", "", fld).lower()
+            match = " MATCH({}, '{}', 'operator=OR;fuzziness=AUTO:1,3;minimum_should_match=30%') ".format(fld, huqie.qieqie(huqie.qie(v)))
-            if v[0] == "%%": v = v[1:-1]
+            replaces.append(("{}{}'{}'".format(r.group(1), r.group(2), r.group(3)), match))
-            match = " MATCH({}, '{}', 'operator=OR;fuzziness=AUTO:1,3;minimum_should_match=30%') ".format(fld, huqie.qie(v))
-            replaces.append((r.group(1)+r.group(2), match))
 
-        for p, r in replaces: sql.replace(p, r)
+        for p, r in replaces: sql = sql.replace(p, r, 1)
+        es_logger.info(f"To es: {sql}")
 
         try:
-            tbl = self.es.sql(sql, fetch_size)
+            tbl = self.es.sql(sql, fetch_size, format)
             return tbl
         except Exception as e:
-            es_logger(f"SQL failure: {sql} =>" + str(e))
+            es_logger.error(f"SQL failure: {sql} =>" + str(e))
 

rag/parser/pdf_parser.py

@@ -53,9 +53,10 @@ class HuParser:
 
     def __remote_call(self, species, images, thr=0.7):
         url = os.environ.get("INFINIFLOW_SERVER")
-        if not url:raise EnvironmentError("Please set environment variable: 'INFINIFLOW_SERVER'")
         token = os.environ.get("INFINIFLOW_TOKEN")
-        if not token:raise EnvironmentError("Please set environment variable: 'INFINIFLOW_TOKEN'")
+        if not url or not token:
+            logging.warning("INFINIFLOW_SERVER is not specified. To maximize the effectiveness, please visit https://github.com/infiniflow/ragflow, and sign in the our demo web site to get token. It's FREE! Using 'export' to set both environment variables: INFINIFLOW_SERVER and INFINIFLOW_TOKEN.")
+            return []
 
         def convert_image_to_bytes(PILimage):
             image = BytesIO()

rag/svr/task_executor.py

@@ -47,7 +47,7 @@ from api.utils.file_utils import get_project_base_directory
 BATCH_SIZE = 64
 
 FACTORY = {
-    ParserType.GENERAL.value: laws,
+    ParserType.GENERAL.value: manual,
     ParserType.PAPER.value: paper,
     ParserType.BOOK.value: book,
     ParserType.PRESENTATION.value: presentation,
@@ -119,8 +119,8 @@ def build(row, cvmdl):
     chunker = FACTORY[row["parser_id"].lower()]
     try:
         cron_logger.info("Chunkking {}/{}".format(row["location"], row["name"]))
-        cks = chunker.chunk(row["name"], MINIO.get(row["kb_id"], row["location"]), row["from_page"], row["to_page"],
+        cks = chunker.chunk(row["name"], binary = MINIO.get(row["kb_id"], row["location"]), from_page=row["from_page"], to_page=row["to_page"],
-                            callback, kb_id=row["kb_id"], parser_config=row["parser_config"])
+                            callback = callback, kb_id=row["kb_id"], parser_config=row["parser_config"])
     except Exception as e:
         if re.search("(No such file|not found)", str(e)):
             callback(-1, "Can not find file <%s>" % row["doc_name"])
@@ -129,7 +129,7 @@ def build(row, cvmdl):
 
         cron_logger.warn("Chunkking {}/{}: {}".format(row["location"], row["name"], str(e)))
 
-        return []
+        return
 
     callback(msg="Finished slicing files. Start to embedding the content.")
 
@@ -211,6 +211,7 @@ def main(comm, mod):
 
         st_tm = timer()
         cks = build(r, cv_mdl)
+        if cks is None:continue
         if not cks:
             tmf.write(str(r["update_time"]) + "\n")
             callback(1., "No chunk! Done!")

rag/utils/es_conn.py

@@ -241,7 +241,7 @@ class HuEs:
         es_logger.error("ES search timeout for 3 times!")
         raise Exception("ES search timeout.")
 
-    def sql(self, sql, fetch_size=128, format="json", timeout=2):
+    def sql(self, sql, fetch_size=128, format="json", timeout="2s"):
         for i in range(3):
             try:
                 res = self.es.sql.query(body={"query": sql, "fetch_size": fetch_size}, format=format, request_timeout=timeout)