diff --git a/docs/examples/hybrid-layer.ipynb b/docs/examples/hybrid-layer.ipynb
index 589e13a1e4d3fc1d924c7bb851c19c56f9f30dd5..d3fb58c5978ea3da709d12eb850d16a846396e0b 100644
--- a/docs/examples/hybrid-layer.ipynb
+++ b/docs/examples/hybrid-layer.ipynb
@@ -151,39 +151,7 @@
"name": "stdout",
"output_type": "stream",
"text": [
- "2067848296 1405\n",
- "2212344012 2520\n",
- "3313717465 206\n",
- "3076736765 769\n",
- "1778150425 4131\n",
- "2067848296 1405\n",
- "202708381 770\n",
- "2212344012 2520\n",
- "3374841595 2375\n",
- "2067848296 1405\n",
- "3508911095 2067\n",
- "3454774732 not in encoder.idx_mapping\n",
- "2379717389 3565\n",
- "298452803 4356\n",
- "1063320047 3369\n",
- "4186256544 713\n",
- "1846246980 858\n",
- "3897916792 643\n",
- "575623047 1476\n",
- "3897916792 643\n"
- ]
- },
- {
- "ename": "ValueError",
- "evalue": "all the input arrays must have same number of dimensions, but the array at index 0 has 2 dimension(s) and the array at index 1 has 1 dimension(s)",
- "output_type": "error",
- "traceback": [
- "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
- "\u001b[0;31mValueError\u001b[0m Traceback (most recent call last)",
- "\u001b[1;32m/Users/jamesbriggs/Documents/projects/aurelio-labs/semantic-router/docs/examples/hybrid-layer.ipynb Cell 14\u001b[0m line \u001b[0;36m3\n\u001b[1;32m <a href='vscode-notebook-cell:/Users/jamesbriggs/Documents/projects/aurelio-labs/semantic-router/docs/examples/hybrid-layer.ipynb#X16sZmlsZQ%3D%3D?line=0'>1</a>\u001b[0m \u001b[39mfrom\u001b[39;00m \u001b[39msemantic_router\u001b[39;00m\u001b[39m.\u001b[39;00m\u001b[39mlayer\u001b[39;00m \u001b[39mimport\u001b[39;00m HybridDecisionLayer\n\u001b[0;32m----> <a href='vscode-notebook-cell:/Users/jamesbriggs/Documents/projects/aurelio-labs/semantic-router/docs/examples/hybrid-layer.ipynb#X16sZmlsZQ%3D%3D?line=2'>3</a>\u001b[0m dl \u001b[39m=\u001b[39m HybridDecisionLayer(encoder\u001b[39m=\u001b[39;49mencoder, decisions\u001b[39m=\u001b[39;49mdecisions)\n",
- "File \u001b[0;32m~/Documents/projects/aurelio-labs/semantic-router/semantic_router/layer.py:137\u001b[0m, in \u001b[0;36mHybridDecisionLayer.__init__\u001b[0;34m(self, encoder, decisions, alpha)\u001b[0m\n\u001b[1;32m 134\u001b[0m \u001b[39mif\u001b[39;00m decisions:\n\u001b[1;32m 135\u001b[0m \u001b[39m# initialize index now\u001b[39;00m\n\u001b[1;32m 136\u001b[0m \u001b[39mfor\u001b[39;00m decision \u001b[39min\u001b[39;00m decisions:\n\u001b[0;32m--> 137\u001b[0m \u001b[39mself\u001b[39;49m\u001b[39m.\u001b[39;49m_add_decision(decision\u001b[39m=\u001b[39;49mdecision)\n",
- "File \u001b[0;32m~/Documents/projects/aurelio-labs/semantic-router/semantic_router/layer.py:156\u001b[0m, in \u001b[0;36mHybridDecisionLayer._add_decision\u001b[0;34m(self, decision)\u001b[0m\n\u001b[1;32m 154\u001b[0m sparse_embeds \u001b[39m=\u001b[39m \u001b[39mself\u001b[39m\u001b[39m.\u001b[39msparse_encoder(decision\u001b[39m.\u001b[39mutterances)\n\u001b[1;32m 155\u001b[0m \u001b[39m# concatenate vectors to create hybrid vecs\u001b[39;00m\n\u001b[0;32m--> 156\u001b[0m embeds \u001b[39m=\u001b[39m np\u001b[39m.\u001b[39;49mconcatenate([\n\u001b[1;32m 157\u001b[0m dense_embeds, sparse_embeds\n\u001b[1;32m 158\u001b[0m ], axis\u001b[39m=\u001b[39;49m\u001b[39m1\u001b[39;49m)\n\u001b[1;32m 160\u001b[0m \u001b[39m# create decision array\u001b[39;00m\n\u001b[1;32m 161\u001b[0m \u001b[39mif\u001b[39;00m \u001b[39mself\u001b[39m\u001b[39m.\u001b[39mcategories \u001b[39mis\u001b[39;00m \u001b[39mNone\u001b[39;00m:\n",
- "\u001b[0;31mValueError\u001b[0m: all the input arrays must have same number of dimensions, but the array at index 0 has 2 dimension(s) and the array at index 1 has 1 dimension(s)"
+ "3454774732 not in encoder.idx_mapping\n"
]
}
],
@@ -195,9 +163,24 @@
},
{
"cell_type": "code",
- "execution_count": null,
+ "execution_count": 6,
"metadata": {},
- "outputs": [],
+ "outputs": [
+ {
+ "ename": "AxisError",
+ "evalue": "axis 1 is out of bounds for array of dimension 1",
+ "output_type": "error",
+ "traceback": [
+ "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
+ "\u001b[0;31mAxisError\u001b[0m Traceback (most recent call last)",
+ "\u001b[1;32m/Users/jamesbriggs/Documents/projects/aurelio-labs/semantic-router/docs/examples/hybrid-layer.ipynb Cell 15\u001b[0m line \u001b[0;36m1\n\u001b[0;32m----> <a href='vscode-notebook-cell:/Users/jamesbriggs/Documents/projects/aurelio-labs/semantic-router/docs/examples/hybrid-layer.ipynb#X20sZmlsZQ%3D%3D?line=0'>1</a>\u001b[0m dl(\u001b[39m\"\u001b[39;49m\u001b[39mdon\u001b[39;49m\u001b[39m'\u001b[39;49m\u001b[39mt you love politics?\u001b[39;49m\u001b[39m\"\u001b[39;49m)\n",
+ "File \u001b[0;32m~/Documents/projects/aurelio-labs/semantic-router/semantic_router/layer.py:141\u001b[0m, in \u001b[0;36mHybridDecisionLayer.__call__\u001b[0;34m(self, text)\u001b[0m\n\u001b[1;32m 140\u001b[0m \u001b[39mdef\u001b[39;00m \u001b[39m__call__\u001b[39m(\u001b[39mself\u001b[39m, text: \u001b[39mstr\u001b[39m) \u001b[39m-\u001b[39m\u001b[39m>\u001b[39m \u001b[39mstr\u001b[39m \u001b[39m|\u001b[39m \u001b[39mNone\u001b[39;00m:\n\u001b[0;32m--> 141\u001b[0m results \u001b[39m=\u001b[39m \u001b[39mself\u001b[39;49m\u001b[39m.\u001b[39;49m_query(text)\n\u001b[1;32m 142\u001b[0m top_class, top_class_scores \u001b[39m=\u001b[39m \u001b[39mself\u001b[39m\u001b[39m.\u001b[39m_semantic_classify(results)\n\u001b[1;32m 143\u001b[0m passed \u001b[39m=\u001b[39m \u001b[39mself\u001b[39m\u001b[39m.\u001b[39m_pass_threshold(top_class_scores, \u001b[39mself\u001b[39m\u001b[39m.\u001b[39mscore_threshold)\n",
+ "File \u001b[0;32m~/Documents/projects/aurelio-labs/semantic-router/semantic_router/layer.py:204\u001b[0m, in \u001b[0;36mHybridDecisionLayer._query\u001b[0;34m(self, text, top_k)\u001b[0m\n\u001b[1;32m 202\u001b[0m sim_d \u001b[39m=\u001b[39m np\u001b[39m.\u001b[39mdot(\u001b[39mself\u001b[39m\u001b[39m.\u001b[39mindex, xq_d\u001b[39m.\u001b[39mT) \u001b[39m/\u001b[39m (index_norm \u001b[39m*\u001b[39m xq_d_norm)\n\u001b[1;32m 203\u001b[0m \u001b[39m# calculate sparse vec similarity\u001b[39;00m\n\u001b[0;32m--> 204\u001b[0m sparse_norm \u001b[39m=\u001b[39m norm(\u001b[39mself\u001b[39;49m\u001b[39m.\u001b[39;49msparse_index, axis\u001b[39m=\u001b[39;49m\u001b[39m1\u001b[39;49m)\n\u001b[1;32m 205\u001b[0m xq_s_norm \u001b[39m=\u001b[39m norm(xq_s\u001b[39m.\u001b[39mT)\n\u001b[1;32m 206\u001b[0m sim_s \u001b[39m=\u001b[39m np\u001b[39m.\u001b[39mdot(\u001b[39mself\u001b[39m\u001b[39m.\u001b[39msparse_index, xq_s\u001b[39m.\u001b[39mT) \u001b[39m/\u001b[39m (sparse_norm \u001b[39m*\u001b[39m xq_s_norm)\n",
+ "File \u001b[0;32m~/opt/anaconda3/envs/decision-layer/lib/python3.11/site-packages/numpy/linalg/linalg.py:2583\u001b[0m, in \u001b[0;36mnorm\u001b[0;34m(x, ord, axis, keepdims)\u001b[0m\n\u001b[1;32m 2580\u001b[0m \u001b[39melif\u001b[39;00m \u001b[39mord\u001b[39m \u001b[39mis\u001b[39;00m \u001b[39mNone\u001b[39;00m \u001b[39mor\u001b[39;00m \u001b[39mord\u001b[39m \u001b[39m==\u001b[39m \u001b[39m2\u001b[39m:\n\u001b[1;32m 2581\u001b[0m \u001b[39m# special case for speedup\u001b[39;00m\n\u001b[1;32m 2582\u001b[0m s \u001b[39m=\u001b[39m (x\u001b[39m.\u001b[39mconj() \u001b[39m*\u001b[39m x)\u001b[39m.\u001b[39mreal\n\u001b[0;32m-> 2583\u001b[0m \u001b[39mreturn\u001b[39;00m sqrt(add\u001b[39m.\u001b[39;49mreduce(s, axis\u001b[39m=\u001b[39;49maxis, keepdims\u001b[39m=\u001b[39;49mkeepdims))\n\u001b[1;32m 2584\u001b[0m \u001b[39m# None of the str-type keywords for ord ('fro', 'nuc')\u001b[39;00m\n\u001b[1;32m 2585\u001b[0m \u001b[39m# are valid for vectors\u001b[39;00m\n\u001b[1;32m 2586\u001b[0m \u001b[39melif\u001b[39;00m \u001b[39misinstance\u001b[39m(\u001b[39mord\u001b[39m, \u001b[39mstr\u001b[39m):\n",
+ "\u001b[0;31mAxisError\u001b[0m: axis 1 is out of bounds for array of dimension 1"
+ ]
+ }
+ ],
"source": [
"dl(\"don't you love politics?\")"
]
@@ -207,10 +190,7 @@
"execution_count": null,
"metadata": {},
"outputs": [],
- "source": [
- "if 3454774732 in encoder.idx_mapping:\n",
- " print(\"yes\")"
- ]
+ "source": []
},
{
"cell_type": "code",
@@ -220,11 +200,13 @@
"source": []
},
{
- "cell_type": "code",
- "execution_count": null,
+ "cell_type": "markdown",
"metadata": {},
- "outputs": [],
- "source": []
+ "source": [
+ "---\n",
+ "\n",
+ "#### Testing"
+ ]
},
{
"cell_type": "code",
@@ -347,6 +329,15 @@
"sparse_vec.shape"
]
},
+ {
+ "cell_type": "markdown",
+ "metadata": {},
+ "source": [
+ "#### Finish Testing\n",
+ "\n",
+ "---"
+ ]
+ },
{
"cell_type": "markdown",
"metadata": {},
diff --git a/semantic_router/encoders/bm25.py b/semantic_router/encoders/bm25.py
index 344f0820ae354645331dc6512456ffe65c60113b..c419ac8ff65c061c4e7f1829d2d23d5aee4b37ea 100644
--- a/semantic_router/encoders/bm25.py
+++ b/semantic_router/encoders/bm25.py
@@ -17,7 +17,7 @@ class BM25Encoder(BaseEncoder):
def __call__(self, docs: list[str]) -> list[list[float]]:
if len(docs) == 1:
- sparse_dicts = self.model.encode_query(docs[0])
+ sparse_dicts = self.model.encode_queries(docs)
elif len(docs) > 1:
sparse_dicts = self.model.encode_documents(docs)
else:
@@ -29,7 +29,6 @@ class BM25Encoder(BaseEncoder):
values = output["values"]
for idx, val in zip(indices, values):
if idx in self.idx_mapping:
- print(idx, self.idx_mapping[idx])
position = self.idx_mapping[idx]
embeds[position] = val
else:
diff --git a/semantic_router/layer.py b/semantic_router/layer.py
index e8b71576c4accd5936621f6da2889411caf2fe97..070c9c0bd37ab3d096e8d85753a93318eb68ce1d 100644
--- a/semantic_router/layer.py
+++ b/semantic_router/layer.py
@@ -111,6 +111,7 @@ class DecisionLayer:
class HybridDecisionLayer:
index = None
+ sparse_index = None
categories = None
score_threshold = 0.82
@@ -150,19 +151,19 @@ class HybridDecisionLayer:
def _add_decision(self, decision: Decision):
# create embeddings
- dense_embeds = self.encoder(decision.utterances) * self.alpha
- sparse_embeds = self.sparse_encoder(decision.utterances) * (1 - self.alpha)
- # concatenate vectors to create hybrid vecs
- embeds = np.concatenate([
- dense_embeds, sparse_embeds
- ], axis=1)
+ dense_embeds = np.array(
+ self.encoder(decision.utterances)
+ ) # * self.alpha
+ sparse_embeds = np.array(
+ self.sparse_encoder(decision.utterances)
+ ) # * (1 - self.alpha)
# create decision array
if self.categories is None:
- self.categories = np.array([decision.name] * len(embeds))
+ self.categories = np.array([decision.name] * len(decision.utterances))
self.utterances = np.array(decision.utterances)
else:
- str_arr = np.array([decision.name] * len(embeds))
+ str_arr = np.array([decision.name] * len(decision.utterances))
self.categories = np.concatenate([self.categories, str_arr])
self.utterances = np.concatenate([
self.utterances,
@@ -170,17 +171,15 @@ class HybridDecisionLayer:
])
# create utterance array (the dense index)
if self.index is None:
- self.index = np.array(dense_embeds)
+ self.index = dense_embeds
else:
- embed_arr = np.array(dense_embeds)
- self.index = np.concatenate([self.index, embed_arr])
+ self.index = np.concatenate([self.index, dense_embeds])
# create sparse utterance array
if self.sparse_index is None:
- self.sparse_index = np.array(sparse_embeds)
+ self.sparse_index = sparse_embeds
else:
- sparse_embed_arr = np.array(sparse_embeds)
self.sparse_index = np.concatenate([
- self.sparse_index, sparse_embed_arr
+ self.sparse_index, sparse_embeds
])
def _query(self, text: str, top_k: int = 5):
@@ -195,17 +194,21 @@ class HybridDecisionLayer:
xq_s = np.squeeze(xq_s)
# convex scaling
xq_d, xq_s = self._convex_scaling(xq_d, xq_s)
- # concatenate to create single hybrid vec
- xq = np.concatenate([xq_d, xq_s], axis=1)
if self.index is not None:
+ # calculate dense vec similarity
index_norm = norm(self.index, axis=1)
- xq_norm = norm(xq.T)
- sim = np.dot(self.index, xq.T) / (index_norm * xq_norm)
+ xq_d_norm = norm(xq_d.T)
+ sim_d = np.dot(self.index, xq_d.T) / (index_norm * xq_d_norm)
+ # calculate sparse vec similarity
+ sparse_norm = norm(self.sparse_index, axis=1)
+ xq_s_norm = norm(xq_s.T)
+ sim_s = np.dot(self.sparse_index, xq_s.T) / (sparse_norm * xq_s_norm)
+ total_sim = (sim_d + sim_s)
# get indices of top_k records
- top_k = min(top_k, sim.shape[0])
- idx = np.argpartition(sim, -top_k)[-top_k:]
- scores = sim[idx]
+ top_k = min(top_k, total_sim.shape[0])
+ idx = np.argpartition(total_sim, -top_k)[-top_k:]
+ scores = total_sim[idx]
# get the utterance categories (decision names)
decisions = self.categories[idx] if self.categories is not None else []
return [
@@ -216,8 +219,8 @@ class HybridDecisionLayer:
def _convex_scaling(self, dense: list[float], sparse: list[float]):
# scale sparse and dense vecs
- dense = dense * self.alpha
- sparse = sparse * (1 - self.alpha)
+ dense = np.array(dense) * self.alpha
+ sparse = np.array(sparse) * (1 - self.alpha)
return dense, sparse
def _semantic_classify(self, query_results: list[dict]) -> tuple[str, list[float]]: