feat: hybrid collection management — agent can create/delete/list collections via prompt, remove RAG_COLLECTIONS config, switch to ONNX MiniLM embedding (local, no API)

config.py

@@ -14,9 +14,4 @@ AGENT_MAX_ITERATIONS  = int( os.getenv("AGENT_MAX_ITERATIONS",  default="10"
 MAX_TOOL_OUTPUT       = int( os.getenv("MAX_TOOL_OUTPUT",       default="4000"                       ) )
 # RAG Configuration
 RAG_PERSIST_DIR       =      os.getenv("RAG_PERSIST_DIR",       default="chroma_db"                  )
-RAG_EMBEDDING_MODEL   =      os.getenv("RAG_EMBEDDING_MODEL",   default="nomic-embed-text"           )
+# Embedding: ChromaDB ONNX default (all-MiniLM-L6-v2, lokal, tidak perlu API call)
-RAG_COLLECTIONS       = {
-    "food_recommendations": {
-        "description": "Menu makanan, preferensi pelanggan, data kuliner"
-    },
-}

hendrik.py

@@ -16,8 +16,10 @@ tools_definition = [
     gadget.tools_mapping( schema = coder.schema_git_operation,    handler = coder.git_operation    ),
     gadget.tools_mapping( schema = rag.schema_store_knowledge,    handler = rag.store_knowledge    ),
     gadget.tools_mapping( schema = rag.schema_search_knowledge,   handler = rag.search_knowledge   ),
-    gadget.tools_mapping( schema = rag.schema_list_collections,   handler = rag.list_collections   ),
+    gadget.tools_mapping( schema = rag.schema_create_collection,    handler = rag.create_collection    ),
-    gadget.tools_mapping( schema = rag.schema_inspect_collection, handler = rag.inspect_collection ),
+    gadget.tools_mapping( schema = rag.schema_delete_collection,    handler = rag.delete_collection    ),
+    gadget.tools_mapping( schema = rag.schema_list_collections,     handler = rag.list_collections     ),
+    gadget.tools_mapping( schema = rag.schema_inspect_collection,   handler = rag.inspect_collection   ),
 ]
 
 # Ekstrak dari tools_definition ke dua format berbeda

scripts/gadget.py

@@ -34,13 +34,16 @@ def build_system_prompt(tools_definition):
         "All file operations are relative to this directory.",
         "",
         "RAG capabilities (knowledge retrieval):",
-        "- list_collections       → see available knowledge bases.",
+        "- list_collections       → see available collections & doc counts.",
+        "- create_collection      → create a new collection for a new topic.",
+        "- delete_collection      → permanently remove a collection and its data.",
         "- inspect_collection     → learn metadata fields before searching.",
         "- search_knowledge       → semantic search + optional metadata filter.",
         "- store_knowledge        → save docs with rich metadata for later use.",
         "",
-        "RAG workflow: inspect → search → reason. Always inspect a collection",
+        "You can create collections yourself! When you encounter a new topic,",
-        "first to discover its metadata keys, then use them in search filters."
+        "use create_collection first, then store_knowledge to populate it.",
+        "Always inspect_collection to discover metadata keys before filtering."
     ])
     return "\n".join(lines)
 

tools/rag.py

@@ -1,42 +1,13 @@
 import json
-import urllib.request
-import urllib.error
-from urllib.parse import urlparse
 
 import chromadb
 from chromadb.config import Settings
 import config
 
 
-# ── Embedding (Ollama) ───────────────────────────────────────────────
-
-from chromadb.api.types import EmbeddingFunction, Embeddings
-
-class OllamaEmbeddingFunction(EmbeddingFunction):
-    def __init__(self, base_url, model):
-        parsed = urlparse(base_url.rstrip('/'))
-        self.ollama_base = f"{parsed.scheme}://{parsed.netloc}"
-        self.model = model
-
-    def __call__(self, input) -> Embeddings:
-        url = f"{self.ollama_base}/api/embed"
-        texts = input if isinstance(input, list) else [input]
-        payload = {"model": self.model, "input": texts}
-        data = json.dumps(payload).encode('utf-8')
-        req = urllib.request.Request(url, data=data, method='POST')
-        req.add_header('Content-Type', 'application/json')
-        try:
-            with urllib.request.urlopen(req, timeout=30) as resp:
-                response = json.loads(resp.read().decode('utf-8'))
-            return response["embeddings"]
-        except Exception as e:
-            raise RuntimeError(f"Embedding error: {e}")
-
-
 # ── ChromaDB singleton ───────────────────────────────────────────────
 
 _store = None
-_ef = None
 
 def _get_store():
     global _store
@@ -47,17 +18,9 @@ def _get_store():
         )
     return _store
 
-def _get_ef():
-    global _ef
-    if _ef is None:
-        _ef = OllamaEmbeddingFunction(config.llm_baseurl, config.RAG_EMBEDDING_MODEL)
-    return _ef
-
 def _collection(name):
-    if name not in config.RAG_COLLECTIONS:
+    """Get or create collection — uses ChromaDB's default ONNX embedding (all-MiniLM-L6-v2)."""
-        avail = ", ".join(config.RAG_COLLECTIONS)
+    return _get_store().get_or_create_collection(name=name)
-        raise ValueError(f"Unknown collection '{name}'. Available: {avail}")
-    return _get_store().get_or_create_collection(name=name, embedding_function=_get_ef())
 
 
 # ── Tool schemas ─────────────────────────────────────────────────────
@@ -139,11 +102,55 @@ schema_search_knowledge = {
     }
 }
 
+schema_create_collection = {
+    "type": "function",
+    "function": {
+        "name": "create_collection",
+        "description": (
+            "Create a new RAG collection for a new topic/domain. Use a short, descriptive name "
+            "with underscores (e.g., 'tanaman_hias', 'customer_profiles'). Optionally provide a description."
+        ),
+        "parameters": {
+            "type": "object",
+            "properties": {
+                "name": {
+                    "type": "string",
+                    "description": "Collection name (lowercase, underscores for spaces)"
+                },
+                "description": {
+                    "type": "string",
+                    "description": "What this collection stores",
+                    "default": ""
+                }
+            },
+            "required": ["name"]
+        }
+    }
+}
+
+schema_delete_collection = {
+    "type": "function",
+    "function": {
+        "name": "delete_collection",
+        "description": "Permanently delete an entire RAG collection and all documents in it. Use with caution — this cannot be undone.",
+        "parameters": {
+            "type": "object",
+            "properties": {
+                "name": {
+                    "type": "string",
+                    "description": "Collection name to delete"
+                }
+            },
+            "required": ["name"]
+        }
+    }
+}
+
 schema_list_collections = {
     "type": "function",
     "function": {
         "name": "list_collections",
-        "description": "List all available RAG collections defined in config with their descriptions.",
+        "description": "List all existing RAG collections with their document count and description.",
         "parameters": {"type": "object", "properties": {}}
     }
 }
@@ -231,13 +238,34 @@ def search_knowledge(collection, query, n_results=5, filter=None):
         return f"Error: {e}"
 
 
+def create_collection(name, description=""):
+    try:
+        col = _get_store().get_or_create_collection(name=name)
+        col.modify(metadata={"description": description})
+        return f"Collection '{name}' is ready."
+    except Exception as e:
+        return f"Error: {e}"
+
+def delete_collection(name):
+    try:
+        _get_store().delete_collection(name)
+        return f"Deleted collection '{name}'."
+    except Exception as e:
+        return f"Error: {e}"
+
 def list_collections():
     try:
-        if not config.RAG_COLLECTIONS:
+        cols = _get_store().list_collections()
-            return "No collections defined in config."
+        if not cols:
-        return "Available collections:\n" + "\n".join(
+            return "No collections exist yet."
-            f"- {n}: {i.get('description', '')}" for n, i in config.RAG_COLLECTIONS.items()
+        out = ["Available collections:"]
-        )
+        for col in cols:
+            meta = col.metadata or {}
+            desc = meta.get("description", "")
+            cnt = col.count()
+            tag = f" ({desc})" if desc else ""
+            out.append(f"- {col.name}{tag} [{cnt} docs]")
+        return "\n".join(out)
     except Exception as e:
         return f"Error: {e}"