Don't select vector on infinity (#11151)

### What problem does this PR solve?

Don't select vector on infinity

### Type of change

- [x] Performance Improvement

api/apps/llm_app.py

@@ -348,7 +348,7 @@ def list_app():
         facts = set([o.to_dict()["llm_factory"] for o in objs if o.api_key and o.status == StatusEnum.VALID.value])
         status = {(o.llm_name + "@" + o.llm_factory) for o in objs if o.status == StatusEnum.VALID.value}
         llms = LLMService.get_all()
-        llms = [m.to_dict() for m in llms if m.status == StatusEnum.VALID.value and m.fid not in weighted and (m.llm_name + "@" + m.fid) in status]
+        llms = [m.to_dict() for m in llms if m.status == StatusEnum.VALID.value and m.fid not in weighted and (m.fid == 'Builtin' or (m.llm_name + "@" + m.fid) in status)]
         for m in llms:
             m["available"] = m["fid"] in facts or m["llm_name"].lower() == "flag-embedding" or m["fid"] in self_deployed
             if "tei-" in os.getenv("COMPOSE_PROFILES", "") and m["model_type"] == LLMType.EMBEDDING and m["fid"] == "Builtin" and m["llm_name"] == os.getenv("TEI_MODEL", ""):

rag/llm/embedding_model.py

@@ -31,6 +31,7 @@ from common.log_utils import log_exception
 from common.token_utils import num_tokens_from_string, truncate
 from common import settings
 import logging
+import base64
 
 
 class Base(ABC):
@@ -377,6 +378,46 @@ class JinaEmbed(Base):
         return np.array(embds[0]), cnt
 
 
+class JinaMultiVecEmbed(Base):
+    _FACTORY_NAME = "Jina"
+
+    def __init__(self, key, model_name="jina-embeddings-v4", base_url="https://api.jina.ai/v1/embeddings"):
+        self.base_url = "https://api.jina.ai/v1/embeddings"
+        self.headers = {"Content-Type": "application/json", "Authorization": f"Bearer {key}"}
+        self.model_name = model_name
+
+    def encode(self, texts: list[str|bytes], task="retrieval.passage"):
+        batch_size = 16
+        ress = []
+        token_count = 0
+        input = []
+        for text in texts:
+            if isinstance(text, str):
+                input.append({"text": text})
+            elif isinstance(text, bytes):
+                img_b64s = None
+                try:
+                    base64.b64decode(text, validate=True)
+                    img_b64s = text.decode('utf8')
+                except Exception:
+                    img_b64s = base64.b64encode(text).decode('utf8')
+                input.append({"image": img_b64s})  # base64 encoded image
+        for i in range(0, len(texts), batch_size):
+            data = {"model": self.model_name, "task": task, "truncate": True, "return_multivector": True, "input": input[i : i + batch_size]}
+            response = requests.post(self.base_url, headers=self.headers, json=data)
+            try:
+                res = response.json()
+                ress.extend([d["embeddings"] for d in res["data"]])
+                token_count += self.total_token_count(res)
+            except Exception as _e:
+                log_exception(_e, response)
+        return np.array(ress), token_count
+
+    def encode_queries(self, text):
+        embds, cnt = self.encode([text], task="retrieval.query")
+        return np.array(embds[0]), cnt
+
+
 class MistralEmbed(Base):
     _FACTORY_NAME = "Mistral"
 

rag/utils/infinity_conn.py

@@ -156,6 +156,7 @@ class InfinityConnection(DocStoreConnection):
             msg = f"Infinity {infinity_uri} is unhealthy in 120s."
             logger.error(msg)
             raise Exception(msg)
+        self.column_vec_patt = re.compile(r"q_(?P<vector_size>\d+)_vec")
         logger.info(f"Infinity {infinity_uri} is healthy.")
 
     def _migrate_db(self, inf_conn):
@@ -323,7 +324,8 @@ class InfinityConnection(DocStoreConnection):
                 output.append(score_func)
             if PAGERANK_FLD not in output:
                 output.append(PAGERANK_FLD)
-        output = [f for f in output if f != "_score"]
+        output = [f for f in output if f != "_score" and self.column_vec_patt.match(f) is None]
+        logger.info(f"infinity output fields: {output}")
         if limit <= 0:
             # ElasticSearch default limit is 10000
             limit = 10000