From 587a80915659dec39ba1376cc09b7209e0151541 Mon Sep 17 00:00:00 2001
From: nogueiraguh <nogueiraguh@gmail.com>
Date: Wed, 11 Mar 2026 03:12:57 -0300
Subject: [PATCH] feat: auto-detect embedding dimensions per table in consult
 endpoint
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- Detect VECTOR_DIMENSION_COUNT per table before querying
- Map dimensions to correct embedding model (1536→small, 3072→large)
- Cache query embeddings by dimension to avoid redundant API calls
- All tables now searchable regardless of which model generated them
---
 backend/app.py | 50 ++++++++++++++++++++++++++++++++++----------------
 1 file changed, 34 insertions(+), 16 deletions(-)

diff --git a/backend/app.py b/backend/app.py
index 758d1bf..ca9f313 100644
--- a/backend/app.py
+++ b/backend/app.py
@@ -2458,6 +2458,22 @@ def _embed_text(text: str, genai_cfg: dict, embedding_model_id: str) -> list:
     response = client.embed_text(embed_detail)
     return response.data.embeddings[0]
 
+_DIM_TO_MODEL = {1536: "openai.text-embedding-3-small", 3072: "openai.text-embedding-3-large"}
+
+def _get_table_embedding_dim(cfg: dict, table_name: str) -> int:
+    """Detect the embedding dimension of a table by sampling one row."""
+    conn = _get_adb_connection(cfg)
+    try:
+        cur = conn.cursor()
+        cur.execute(f'SELECT VECTOR_DIMENSION_COUNT(EMBEDDING) FROM "{table_name}" FETCH FIRST 1 ROWS ONLY')
+        row = cur.fetchone()
+        cur.close()
+        return int(row[0]) if row and row[0] else 0
+    except Exception:
+        return 0
+    finally:
+        conn.close()
+
 def _vector_search(cfg: dict, query_embedding: list, top_k: int = 5, table_name: str = None, tenancy: str = None) -> list:
     """Search ADB vector store using cosine similarity. Returns top-K documents.
     If tenancy is provided, filters METADATA JSON to match that tenancy only."""
@@ -3335,35 +3351,37 @@ async def consult_embeddings(req: ConsultQuery, u=Depends(current_user)):
     # Collect results from all active ADB configs + tables
     all_docs = []
     for adb_cfg in adb_configs:
-        log.info(f"Consult: processing ADB {adb_cfg['id']} ({adb_cfg.get('config_name','')})")
         try:
             emb_genai = _resolve_embed_config(oci_config_id=adb_cfg.get("oci_config_id"))
-            log.info(f"Consult: resolved embed config, compartment={emb_genai.get('compartment_id','?')[:30]}, region={emb_genai.get('genai_region','?')}")
         except Exception as e:
             log.warning(f"Consult: resolve config failed for {adb_cfg['id']}: {e}")
             continue
-        emb_model = adb_cfg.get("embedding_model_id", "")
-        if not emb_model:
-            log.warning(f"Consult: no embedding_model_id for {adb_cfg['id']}")
-            continue
-        try:
-            query_embedding = _embed_text(req.query, emb_genai, emb_model)
-            log.info(f"Consult: embedded query, dims={len(query_embedding)}")
-        except Exception as e:
-            log.warning(f"Consult: embed failed for {adb_cfg['id']}: {e}")
-            continue
         tables = _get_tables_for_config(adb_cfg["id"], active_only=True)
-        log.info(f"Consult: {len(tables)} active tables, filter='{req.table_name}'")
         if req.table_name:
             tables = [t for t in tables if t["table_name"] == req.table_name]
+        # Group tables by embedding dimension and generate one query embedding per dimension
+        embeddings_cache = {}  # dim -> query_embedding
         for tbl in tables:
+            tbl_name = tbl["table_name"]
             try:
-                docs = _vector_search(adb_cfg, query_embedding, top_k=req.top_k, table_name=tbl["table_name"])
+                dim = _get_table_embedding_dim(adb_cfg, tbl_name)
+                if not dim:
+                    log.warning(f"Consult: table {tbl_name} is empty or has no embeddings")
+                    continue
+                if dim not in embeddings_cache:
+                    model = _DIM_TO_MODEL.get(dim, adb_cfg.get("embedding_model_id", ""))
+                    if not model:
+                        log.warning(f"Consult: no model for dim={dim} on {tbl_name}")
+                        continue
+                    embeddings_cache[dim] = _embed_text(req.query, emb_genai, model)
+                    log.info(f"Consult: embedded query for dim={dim} using {model}")
+                query_embedding = embeddings_cache[dim]
+                docs = _vector_search(adb_cfg, query_embedding, top_k=req.top_k, table_name=tbl_name)
                 for d in docs:
-                    d["source"] = f"{tbl['table_name']}"
+                    d["source"] = tbl_name
                 all_docs.extend(docs)
             except Exception as e:
-                log.warning(f"Consult: search failed on {tbl['table_name']}: {type(e).__name__}: {e}")
+                log.warning(f"Consult: failed on {tbl_name}: {type(e).__name__}: {str(e)[:200]}")
     if not all_docs:
         return {"answer": "Nenhum resultado encontrado nas bases vetoriais.", "documents": [], "total": 0}
     # Sort by distance and take top results