diff --git a/coverage.xml b/coverage.xml
index 3c9c2e7c7acc2890d77de9dd322f29415a4d39aa..612cac462cea738d5f3a86626f804df04a1c6141 100644
--- a/coverage.xml
+++ b/coverage.xml
@@ -1,5 +1,5 @@
<?xml version="1.0" ?>
-<coverage version="7.3.2" timestamp="1702458856280" lines-valid="311" lines-covered="311" line-rate="1" branches-covered="0" branches-valid="0" branch-rate="0" complexity="0">
+<coverage version="7.3.2" timestamp="1702461637601" lines-valid="315" lines-covered="315" line-rate="1" branches-covered="0" branches-valid="0" branch-rate="0" complexity="0">
<!-- Generated by coverage.py: https://coverage.readthedocs.io/en/7.3.2 -->
<!-- Based on https://raw.githubusercontent.com/cobertura/web/master/htdocs/xml/coverage-04.dtd -->
<sources>
@@ -12,10 +12,11 @@
<methods/>
<lines>
<line number="1" hits="1"/>
- <line number="3" hits="1"/>
+ <line number="2" hits="1"/>
+ <line number="4" hits="1"/>
</lines>
</class>
- <class name="layer.py" filename="layer.py" complexity="0" line-rate="1" branch-rate="0">
+ <class name="hybrid_layer.py" filename="hybrid_layer.py" complexity="0" line-rate="1" branch-rate="0">
<methods/>
<lines>
<line number="1" hits="1"/>
@@ -23,149 +24,157 @@
<line number="3" hits="1"/>
<line number="5" hits="1"/>
<line number="11" hits="1"/>
- <line number="12" hits="1"/>
+ <line number="14" hits="1"/>
<line number="15" hits="1"/>
<line number="16" hits="1"/>
<line number="17" hits="1"/>
<line number="18" hits="1"/>
<line number="20" hits="1"/>
- <line number="21" hits="1"/>
<line number="23" hits="1"/>
<line number="24" hits="1"/>
<line number="25" hits="1"/>
- <line number="26" hits="1"/>
+ <line number="27" hits="1"/>
<line number="28" hits="1"/>
+ <line number="29" hits="1"/>
<line number="30" hits="1"/>
<line number="32" hits="1"/>
<line number="34" hits="1"/>
- <line number="35" hits="1"/>
<line number="36" hits="1"/>
<line number="37" hits="1"/>
- <line number="38" hits="1"/>
<line number="39" hits="1"/>
+ <line number="40" hits="1"/>
<line number="41" hits="1"/>
+ <line number="42" hits="1"/>
+ <line number="43" hits="1"/>
+ <line number="44" hits="1"/>
<line number="46" hits="1"/>
<line number="48" hits="1"/>
+ <line number="49" hits="1"/>
<line number="51" hits="1"/>
- <line number="52" hits="1"/>
+ <line number="53" hits="1"/>
<line number="54" hits="1"/>
- <line number="55" hits="1"/>
- <line number="57" hits="1"/>
- <line number="58" hits="1"/>
+ <line number="59" hits="1"/>
<line number="60" hits="1"/>
<line number="61" hits="1"/>
<line number="63" hits="1"/>
+ <line number="64" hits="1"/>
<line number="65" hits="1"/>
- <line number="68" hits="1"/>
- <line number="71" hits="1"/>
+ <line number="69" hits="1"/>
+ <line number="70" hits="1"/>
+ <line number="72" hits="1"/>
<line number="74" hits="1"/>
<line number="75" hits="1"/>
- <line number="82" hits="1"/>
- <line number="83" hits="1"/>
- <line number="89" hits="1"/>
+ <line number="77" hits="1"/>
+ <line number="79" hits="1"/>
+ <line number="84" hits="1"/>
+ <line number="85" hits="1"/>
+ <line number="87" hits="1"/>
+ <line number="88" hits="1"/>
+ <line number="90" hits="1"/>
+ <line number="92" hits="1"/>
<line number="94" hits="1"/>
<line number="95" hits="1"/>
- <line number="97" hits="1"/>
+ <line number="96" hits="1"/>
+ <line number="98" hits="1"/>
<line number="99" hits="1"/>
<line number="100" hits="1"/>
- <line number="102" hits="1"/>
+ <line number="101" hits="1"/>
<line number="103" hits="1"/>
+ <line number="104" hits="1"/>
+ <line number="105" hits="1"/>
<line number="107" hits="1"/>
- <line number="109" hits="1"/>
+ <line number="108" hits="1"/>
<line number="110" hits="1"/>
- <line number="111" hits="1"/>
<line number="112" hits="1"/>
- <line number="113" hits="1"/>
<line number="114" hits="1"/>
<line number="115" hits="1"/>
- <line number="117" hits="1"/>
+ <line number="116" hits="1"/>
+ <line number="118" hits="1"/>
+ <line number="119" hits="1"/>
<line number="120" hits="1"/>
+ <line number="121" hits="1"/>
+ <line number="122" hits="1"/>
<line number="123" hits="1"/>
+ <line number="124" hits="1"/>
<line number="126" hits="1"/>
- <line number="128" hits="1"/>
<line number="129" hits="1"/>
<line number="130" hits="1"/>
- <line number="132" hits="1"/>
+ <line number="133" hits="1"/>
<line number="135" hits="1"/>
<line number="136" hits="1"/>
<line number="137" hits="1"/>
- <line number="138" hits="1"/>
<line number="139" hits="1"/>
- <line number="141" hits="1"/>
- <line number="144" hits="1"/>
- <line number="145" hits="1"/>
- <line number="146" hits="1"/>
- <line number="148" hits="1"/>
- <line number="149" hits="1"/>
- <line number="150" hits="1"/>
- <line number="151" hits="1"/>
- <line number="153" hits="1"/>
- <line number="155" hits="1"/>
- <line number="157" hits="1"/>
- <line number="158" hits="1"/>
- <line number="160" hits="1"/>
- <line number="161" hits="1"/>
- <line number="162" hits="1"/>
- <line number="163" hits="1"/>
- <line number="164" hits="1"/>
- <line number="165" hits="1"/>
- <line number="167" hits="1"/>
- <line number="169" hits="1"/>
- <line number="170" hits="1"/>
- <line number="172" hits="1"/>
- <line number="174" hits="1"/>
- <line number="175" hits="1"/>
- <line number="180" hits="1"/>
- <line number="181" hits="1"/>
- <line number="182" hits="1"/>
- <line number="184" hits="1"/>
- <line number="185" hits="1"/>
- <line number="186" hits="1"/>
- <line number="190" hits="1"/>
- <line number="191" hits="1"/>
- <line number="193" hits="1"/>
- <line number="195" hits="1"/>
- <line number="196" hits="1"/>
- <line number="198" hits="1"/>
- <line number="200" hits="1"/>
- <line number="205" hits="1"/>
- <line number="206" hits="1"/>
- <line number="208" hits="1"/>
- <line number="209" hits="1"/>
- <line number="211" hits="1"/>
- <line number="213" hits="1"/>
- <line number="215" hits="1"/>
- <line number="216" hits="1"/>
- <line number="217" hits="1"/>
- <line number="219" hits="1"/>
- <line number="220" hits="1"/>
- <line number="221" hits="1"/>
- <line number="222" hits="1"/>
- <line number="224" hits="1"/>
- <line number="225" hits="1"/>
- <line number="226" hits="1"/>
- <line number="228" hits="1"/>
- <line number="229" hits="1"/>
- <line number="233" hits="1"/>
- <line number="235" hits="1"/>
- <line number="237" hits="1"/>
- <line number="238" hits="1"/>
- <line number="239" hits="1"/>
- <line number="241" hits="1"/>
- <line number="242" hits="1"/>
- <line number="243" hits="1"/>
- <line number="244" hits="1"/>
- <line number="245" hits="1"/>
- <line number="246" hits="1"/>
- <line number="247" hits="1"/>
- <line number="249" hits="1"/>
- <line number="252" hits="1"/>
- <line number="255" hits="1"/>
- <line number="258" hits="1"/>
- <line number="260" hits="1"/>
- <line number="261" hits="1"/>
- <line number="262" hits="1"/>
- <line number="264" hits="1"/>
+ </lines>
+ </class>
+ <class name="layer.py" filename="layer.py" complexity="0" line-rate="1" branch-rate="0">
+ <methods/>
+ <lines>
+ <line number="1" hits="1"/>
+ <line number="3" hits="1"/>
+ <line number="8" hits="1"/>
+ <line number="9" hits="1"/>
+ <line number="12" hits="1"/>
+ <line number="13" hits="1"/>
+ <line number="14" hits="1"/>
+ <line number="15" hits="1"/>
+ <line number="17" hits="1"/>
+ <line number="18" hits="1"/>
+ <line number="20" hits="1"/>
+ <line number="21" hits="1"/>
+ <line number="22" hits="1"/>
+ <line number="23" hits="1"/>
+ <line number="25" hits="1"/>
+ <line number="27" hits="1"/>
+ <line number="29" hits="1"/>
+ <line number="31" hits="1"/>
+ <line number="32" hits="1"/>
+ <line number="33" hits="1"/>
+ <line number="34" hits="1"/>
+ <line number="35" hits="1"/>
+ <line number="36" hits="1"/>
+ <line number="38" hits="1"/>
+ <line number="40" hits="1"/>
+ <line number="42" hits="1"/>
+ <line number="45" hits="1"/>
+ <line number="46" hits="1"/>
+ <line number="48" hits="1"/>
+ <line number="49" hits="1"/>
+ <line number="51" hits="1"/>
+ <line number="52" hits="1"/>
+ <line number="54" hits="1"/>
+ <line number="55" hits="1"/>
+ <line number="57" hits="1"/>
+ <line number="59" hits="1"/>
+ <line number="62" hits="1"/>
+ <line number="65" hits="1"/>
+ <line number="66" hits="1"/>
+ <line number="67" hits="1"/>
+ <line number="74" hits="1"/>
+ <line number="75" hits="1"/>
+ <line number="81" hits="1"/>
+ <line number="86" hits="1"/>
+ <line number="87" hits="1"/>
+ <line number="89" hits="1"/>
+ <line number="91" hits="1"/>
+ <line number="92" hits="1"/>
+ <line number="94" hits="1"/>
+ <line number="95" hits="1"/>
+ <line number="97" hits="1"/>
+ <line number="99" hits="1"/>
+ <line number="100" hits="1"/>
+ <line number="101" hits="1"/>
+ <line number="102" hits="1"/>
+ <line number="103" hits="1"/>
+ <line number="104" hits="1"/>
+ <line number="105" hits="1"/>
+ <line number="107" hits="1"/>
+ <line number="110" hits="1"/>
+ <line number="111" hits="1"/>
+ <line number="114" hits="1"/>
+ <line number="116" hits="1"/>
+ <line number="117" hits="1"/>
+ <line number="118" hits="1"/>
+ <line number="120" hits="1"/>
</lines>
</class>
<class name="linear.py" filename="linear.py" complexity="0" line-rate="1" branch-rate="0">
diff --git a/docs/examples/hybrid-layer.ipynb b/docs/examples/hybrid-layer.ipynb
index 98fccf6972026d62e6d53a3c96492ca90287207e..8b1da5ae75f0a8a9572996b8e416a282d2c48f1b 100644
--- a/docs/examples/hybrid-layer.ipynb
+++ b/docs/examples/hybrid-layer.ipynb
@@ -11,7 +11,7 @@
"cell_type": "markdown",
"metadata": {},
"source": [
- "The Hybrid Layer in the Semantic Router library can improve decision making performance particularly for niche use-cases that contain specific terminology, such as finance or medical. It helps us provide more importance to decision making based on the keywords contained in our utterances and user queries."
+ "The Hybrid Layer in the Semantic Router library can improve making performance particularly for niche use-cases that contain specific terminology, such as finance or medical. It helps us provide more importance to making based on the keywords contained in our utterances and user queries."
]
},
{
@@ -34,36 +34,37 @@
"metadata": {},
"outputs": [],
"source": [
- "!pip install -qU semantic-router==0.0.5"
+ "!pip install -qU semantic-router==0.0.6"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
- "We start by defining a dictionary mapping decisions to example phrases that should trigger those decisions."
+ "We start by defining a dictionary mapping s to example phrases that should trigger those s."
]
},
{
"cell_type": "code",
- "execution_count": null,
- "metadata": {},
- "outputs": [],
- "source": [
- "import os\n",
+ "execution_count": 8,
+ "metadata": {},
+ "outputs": [
+ {
+ "ename": "ImportError",
+ "evalue": "cannot import name 'Route' from 'semantic_router.schema' (/Users/jakit/customers/aurelio/semantic-router/.venv/lib/python3.11/site-packages/semantic_router/schema.py)",
+ "output_type": "error",
+ "traceback": [
+ "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
+ "\u001b[0;31mImportError\u001b[0m Traceback (most recent call last)",
+ "\u001b[1;32m/Users/jakit/customers/aurelio/semantic-router/docs/examples/hybrid-layer.ipynb Cell 7\u001b[0m line \u001b[0;36m1\n\u001b[0;32m----> <a href='vscode-notebook-cell:/Users/jakit/customers/aurelio/semantic-router/docs/examples/hybrid-layer.ipynb#X10sZmlsZQ%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[39mschema\u001b[39;00m \u001b[39mimport\u001b[39;00m Route\n\u001b[1;32m <a href='vscode-notebook-cell:/Users/jakit/customers/aurelio/semantic-router/docs/examples/hybrid-layer.ipynb#X10sZmlsZQ%3D%3D?line=2'>3</a>\u001b[0m politics \u001b[39m=\u001b[39m Route(\n\u001b[1;32m <a href='vscode-notebook-cell:/Users/jakit/customers/aurelio/semantic-router/docs/examples/hybrid-layer.ipynb#X10sZmlsZQ%3D%3D?line=3'>4</a>\u001b[0m name\u001b[39m=\u001b[39m\u001b[39m\"\u001b[39m\u001b[39mpolitics\u001b[39m\u001b[39m\"\u001b[39m,\n\u001b[1;32m <a href='vscode-notebook-cell:/Users/jakit/customers/aurelio/semantic-router/docs/examples/hybrid-layer.ipynb#X10sZmlsZQ%3D%3D?line=4'>5</a>\u001b[0m utterances\u001b[39m=\u001b[39m[\n\u001b[0;32m (...)\u001b[0m\n\u001b[1;32m <a href='vscode-notebook-cell:/Users/jakit/customers/aurelio/semantic-router/docs/examples/hybrid-layer.ipynb#X10sZmlsZQ%3D%3D?line=11'>12</a>\u001b[0m ],\n\u001b[1;32m <a href='vscode-notebook-cell:/Users/jakit/customers/aurelio/semantic-router/docs/examples/hybrid-layer.ipynb#X10sZmlsZQ%3D%3D?line=12'>13</a>\u001b[0m )\n",
+ "\u001b[0;31mImportError\u001b[0m: cannot import name 'Route' from 'semantic_router.schema' (/Users/jakit/customers/aurelio/semantic-router/.venv/lib/python3.11/site-packages/semantic_router/schema.py)"
+ ]
+ }
+ ],
+ "source": [
+ "from semantic_router.schema import Route\n",
"\n",
- "os.environ[\"COHERE_API_KEY\"] = \"BQBiUqqjDRsYl1QKKux4JsqKdDkjyInS5T3Z3eJP\""
- ]
- },
- {
- "cell_type": "code",
- "execution_count": null,
- "metadata": {},
- "outputs": [],
- "source": [
- "from semantic_router.schema import Decision\n",
- "\n",
- "politics = Decision(\n",
+ "politics = Route(\n",
" name=\"politics\",\n",
" utterances=[\n",
" \"isn't politics the best thing ever\",\n",
@@ -89,7 +90,7 @@
"metadata": {},
"outputs": [],
"source": [
- "chitchat = Decision(\n",
+ "chitchat = Route(\n",
" name=\"chitchat\",\n",
" utterances=[\n",
" \"how's the weather today?\",\n",
@@ -100,7 +101,7 @@
" ],\n",
")\n",
"\n",
- "chitchat = Decision(\n",
+ "chitchat = Route(\n",
" name=\"chitchat\",\n",
" utterances=[\n",
" \"how's the weather today?\",\n",
@@ -111,7 +112,7 @@
" ],\n",
")\n",
"\n",
- "decisions = [politics, chitchat]"
+ "routes = [politics, chitchat]"
]
},
{
@@ -127,6 +128,7 @@
"metadata": {},
"outputs": [],
"source": [
+ "import os\n",
"from semantic_router.encoders import CohereEncoder\n",
"from getpass import getpass\n",
"\n",
@@ -141,7 +143,7 @@
"cell_type": "markdown",
"metadata": {},
"source": [
- "Now we define the `DecisionLayer`. When called, the decision layer will consume text (a query) and output the category (`Decision`) it belongs to — to initialize a `DecisionLayer` we need our `encoder` model and a list of `decisions`."
+ "Now we define the `RouteLayer`. When called, the route layer will consume text (a query) and output the category (`Route`) it belongs to — to initialize a `RouteLayer` we need our `encoder` model and a list of `routes`."
]
},
{
@@ -150,9 +152,9 @@
"metadata": {},
"outputs": [],
"source": [
- "from semantic_router.layer import HybridDecisionLayer\n",
+ "from semantic_router.hybrid_layer import HybridRouteLayer\n",
"\n",
- "dl = HybridDecisionLayer(encoder=encoder, decisions=decisions)"
+ "dl = HybridRouteLayer(encoder=encoder, routes=routes)"
]
},
{
@@ -197,7 +199,7 @@
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
- "version": "3.11.5"
+ "version": "3.11.3"
}
},
"nbformat": 4,
diff --git a/semantic_router/__init__.py b/semantic_router/__init__.py
index b10cc687b0924faf516432e89f87bc1f95bb0c9e..0c445bea3ff4efd8f3aa8950e2c772277d93b20c 100644
--- a/semantic_router/__init__.py
+++ b/semantic_router/__init__.py
@@ -1,3 +1,4 @@
-from .layer import RouteLayer, HybridRouteLayer
+from .hybrid_layer import HybridRouteLayer
+from .layer import RouteLayer
__all__ = ["RouteLayer", "HybridRouteLayer"]
diff --git a/semantic_router/hybrid_layer.py b/semantic_router/hybrid_layer.py
new file mode 100644
index 0000000000000000000000000000000000000000..3e9508d6a50862cc523d80e802716f71adeee99d
--- /dev/null
+++ b/semantic_router/hybrid_layer.py
@@ -0,0 +1,139 @@
+import numpy as np
+from numpy.linalg import norm
+from tqdm.auto import tqdm
+
+from semantic_router.encoders import (
+ BaseEncoder,
+ BM25Encoder,
+ CohereEncoder,
+ OpenAIEncoder,
+)
+from semantic_router.schema import Route
+
+
+class HybridRouteLayer:
+ index = None
+ sparse_index = None
+ categories = None
+ score_threshold = 0.82
+
+ def __init__(
+ self, encoder: BaseEncoder, routes: list[Route] = [], alpha: float = 0.3
+ ):
+ self.encoder = encoder
+ self.sparse_encoder = BM25Encoder()
+ self.alpha = alpha
+ # decide on default threshold based on encoder
+ if isinstance(encoder, OpenAIEncoder):
+ self.score_threshold = 0.82
+ elif isinstance(encoder, CohereEncoder):
+ self.score_threshold = 0.3
+ else:
+ self.score_threshold = 0.82
+ # if routes list has been passed, we initialize index now
+ if routes:
+ # initialize index now
+ for route in tqdm(routes):
+ self._add_route(route=route)
+
+ def __call__(self, text: str) -> str | None:
+ results = self._query(text)
+ top_class, top_class_scores = self._semantic_classify(results)
+ passed = self._pass_threshold(top_class_scores, self.score_threshold)
+ if passed:
+ return top_class
+ else:
+ return None
+
+ def add(self, route: Route):
+ self._add_route(route=route)
+
+ def _add_route(self, route: Route):
+ # create embeddings
+ dense_embeds = np.array(self.encoder(route.utterances)) # * self.alpha
+ sparse_embeds = np.array(
+ self.sparse_encoder(route.utterances)
+ ) # * (1 - self.alpha)
+
+ # create route array
+ if self.categories is None:
+ self.categories = np.array([route.name] * len(route.utterances))
+ self.utterances = np.array(route.utterances)
+ else:
+ str_arr = np.array([route.name] * len(route.utterances))
+ self.categories = np.concatenate([self.categories, str_arr])
+ self.utterances = np.concatenate(
+ [self.utterances, np.array(route.utterances)]
+ )
+ # create utterance array (the dense index)
+ if self.index is None:
+ self.index = dense_embeds
+ else:
+ self.index = np.concatenate([self.index, dense_embeds])
+ # create sparse utterance array
+ if self.sparse_index is None:
+ self.sparse_index = sparse_embeds
+ else:
+ self.sparse_index = np.concatenate([self.sparse_index, sparse_embeds])
+
+ def _query(self, text: str, top_k: int = 5):
+ """Given some text, encodes and searches the index vector space to
+ retrieve the top_k most similar records.
+ """
+ # create dense query vector
+ xq_d = np.array(self.encoder([text]))
+ xq_d = np.squeeze(xq_d) # Reduce to 1d array.
+ # create sparse query vector
+ xq_s = np.array(self.sparse_encoder([text]))
+ xq_s = np.squeeze(xq_s)
+ # convex scaling
+ xq_d, xq_s = self._convex_scaling(xq_d, xq_s)
+
+ if self.index is not None:
+ # calculate dense vec similarity
+ index_norm = norm(self.index, axis=1)
+ 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, total_sim.shape[0])
+ idx = np.argpartition(total_sim, -top_k)[-top_k:]
+ scores = total_sim[idx]
+ # get the utterance categories (route names)
+ routes = self.categories[idx] if self.categories is not None else []
+ return [{"route": d, "score": s.item()} for d, s in zip(routes, scores)]
+ else:
+ return []
+
+ def _convex_scaling(self, dense: list[float], sparse: list[float]):
+ # scale sparse and dense vecs
+ 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]]:
+ scores_by_class = {}
+ for result in query_results:
+ score = result["score"]
+ route = result["route"]
+ if route in scores_by_class:
+ scores_by_class[route].append(score)
+ else:
+ scores_by_class[route] = [score]
+
+ # Calculate total score for each class
+ total_scores = {route: sum(scores) for route, scores in scores_by_class.items()}
+ top_class = max(total_scores, key=lambda x: total_scores[x], default=None)
+
+ # Return the top class and its associated scores
+ return str(top_class), scores_by_class.get(top_class, [])
+
+ def _pass_threshold(self, scores: list[float], threshold: float) -> bool:
+ if scores:
+ return max(scores) > threshold
+ else:
+ return False
diff --git a/semantic_router/layer.py b/semantic_router/layer.py
index 591e8f08f60108adc727a309e4d22aa3415fb640..efa4862d1766e4bd677101373e71d8be95897c38 100644
--- a/semantic_router/layer.py
+++ b/semantic_router/layer.py
@@ -1,10 +1,7 @@
import numpy as np
-from numpy.linalg import norm
-from tqdm.auto import tqdm
from semantic_router.encoders import (
BaseEncoder,
- BM25Encoder,
CohereEncoder,
OpenAIEncoder,
)
@@ -121,131 +118,3 @@ class RouteLayer:
return max(scores) > threshold
else:
return False
-
-
-class HybridRouteLayer:
- index = None
- sparse_index = None
- categories = None
- score_threshold = 0.82
-
- def __init__(
- self, encoder: BaseEncoder, routes: list[Route] = [], alpha: float = 0.3
- ):
- self.encoder = encoder
- self.sparse_encoder = BM25Encoder()
- self.alpha = alpha
- # decide on default threshold based on encoder
- if isinstance(encoder, OpenAIEncoder):
- self.score_threshold = 0.82
- elif isinstance(encoder, CohereEncoder):
- self.score_threshold = 0.3
- else:
- self.score_threshold = 0.82
- # if routes list has been passed, we initialize index now
- if routes:
- # initialize index now
- for route in tqdm(routes):
- self._add_route(route=route)
-
- def __call__(self, text: str) -> str | None:
- results = self._query(text)
- top_class, top_class_scores = self._semantic_classify(results)
- passed = self._pass_threshold(top_class_scores, self.score_threshold)
- if passed:
- return top_class
- else:
- return None
-
- def add(self, route: Route):
- self._add_route(route=route)
-
- def _add_route(self, route: Route):
- # create embeddings
- dense_embeds = np.array(self.encoder(route.utterances)) # * self.alpha
- sparse_embeds = np.array(
- self.sparse_encoder(route.utterances)
- ) # * (1 - self.alpha)
-
- # create route array
- if self.categories is None:
- self.categories = np.array([route.name] * len(route.utterances))
- self.utterances = np.array(route.utterances)
- else:
- str_arr = np.array([route.name] * len(route.utterances))
- self.categories = np.concatenate([self.categories, str_arr])
- self.utterances = np.concatenate(
- [self.utterances, np.array(route.utterances)]
- )
- # create utterance array (the dense index)
- if self.index is None:
- self.index = dense_embeds
- else:
- self.index = np.concatenate([self.index, dense_embeds])
- # create sparse utterance array
- if self.sparse_index is None:
- self.sparse_index = sparse_embeds
- else:
- self.sparse_index = np.concatenate([self.sparse_index, sparse_embeds])
-
- def _query(self, text: str, top_k: int = 5):
- """Given some text, encodes and searches the index vector space to
- retrieve the top_k most similar records.
- """
- # create dense query vector
- xq_d = np.array(self.encoder([text]))
- xq_d = np.squeeze(xq_d) # Reduce to 1d array.
- # create sparse query vector
- xq_s = np.array(self.sparse_encoder([text]))
- xq_s = np.squeeze(xq_s)
- # convex scaling
- xq_d, xq_s = self._convex_scaling(xq_d, xq_s)
-
- if self.index is not None:
- # calculate dense vec similarity
- index_norm = norm(self.index, axis=1)
- 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, total_sim.shape[0])
- idx = np.argpartition(total_sim, -top_k)[-top_k:]
- scores = total_sim[idx]
- # get the utterance categories (route names)
- routes = self.categories[idx] if self.categories is not None else []
- return [{"route": d, "score": s.item()} for d, s in zip(routes, scores)]
- else:
- return []
-
- def _convex_scaling(self, dense: list[float], sparse: list[float]):
- # scale sparse and dense vecs
- 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]]:
- scores_by_class = {}
- for result in query_results:
- score = result["score"]
- route = result["route"]
- if route in scores_by_class:
- scores_by_class[route].append(score)
- else:
- scores_by_class[route] = [score]
-
- # Calculate total score for each class
- total_scores = {route: sum(scores) for route, scores in scores_by_class.items()}
- top_class = max(total_scores, key=lambda x: total_scores[x], default=None)
-
- # Return the top class and its associated scores
- return str(top_class), scores_by_class.get(top_class, [])
-
- def _pass_threshold(self, scores: list[float], threshold: float) -> bool:
- if scores:
- return max(scores) > threshold
- else:
- return False
diff --git a/tests/unit/test_hybrid_layer.py b/tests/unit/test_hybrid_layer.py
new file mode 100644
index 0000000000000000000000000000000000000000..94720cd8d9567b19e78a016a03c2cc90b8f62d40
--- /dev/null
+++ b/tests/unit/test_hybrid_layer.py
@@ -0,0 +1,118 @@
+import pytest
+
+from semantic_router.encoders import BaseEncoder, CohereEncoder, OpenAIEncoder
+from semantic_router.hybrid_layer import HybridRouteLayer
+from semantic_router.schema import Route
+
+
+def mock_encoder_call(utterances):
+ # Define a mapping of utterances to return values
+ mock_responses = {
+ "Hello": [0.1, 0.2, 0.3],
+ "Hi": [0.4, 0.5, 0.6],
+ "Goodbye": [0.7, 0.8, 0.9],
+ "Bye": [1.0, 1.1, 1.2],
+ "Au revoir": [1.3, 1.4, 1.5],
+ }
+ return [mock_responses.get(u, [0, 0, 0]) for u in utterances]
+
+
+@pytest.fixture
+def base_encoder():
+ return BaseEncoder(name="test-encoder")
+
+
+@pytest.fixture
+def cohere_encoder(mocker):
+ mocker.patch.object(CohereEncoder, "__call__", side_effect=mock_encoder_call)
+ return CohereEncoder(name="test-cohere-encoder", cohere_api_key="test_api_key")
+
+
+@pytest.fixture
+def openai_encoder(mocker):
+ mocker.patch.object(OpenAIEncoder, "__call__", side_effect=mock_encoder_call)
+ return OpenAIEncoder(name="test-openai-encoder", openai_api_key="test_api_key")
+
+
+@pytest.fixture
+def routes():
+ return [
+ Route(name="Route 1", utterances=["Hello", "Hi"]),
+ Route(name="Route 2", utterances=["Goodbye", "Bye", "Au revoir"]),
+ ]
+
+
+class TestHybridRouteLayer:
+ def test_initialization(self, openai_encoder, routes):
+ route_layer = HybridRouteLayer(encoder=openai_encoder, routes=routes)
+ assert route_layer.index is not None and route_layer.categories is not None
+ assert route_layer.score_threshold == 0.82
+ assert len(route_layer.index) == 5
+ assert len(set(route_layer.categories)) == 2
+
+ def test_initialization_different_encoders(self, cohere_encoder, openai_encoder):
+ route_layer_cohere = HybridRouteLayer(encoder=cohere_encoder)
+ assert route_layer_cohere.score_threshold == 0.3
+
+ route_layer_openai = HybridRouteLayer(encoder=openai_encoder)
+ assert route_layer_openai.score_threshold == 0.82
+
+ def test_add_route(self, openai_encoder):
+ route_layer = HybridRouteLayer(encoder=openai_encoder)
+ route = Route(name="Route 3", utterances=["Yes", "No"])
+ route_layer.add(route)
+ assert route_layer.index is not None and route_layer.categories is not None
+ assert len(route_layer.index) == 2
+ assert len(set(route_layer.categories)) == 1
+
+ def test_add_multiple_routes(self, openai_encoder, routes):
+ route_layer = HybridRouteLayer(encoder=openai_encoder)
+ for route in routes:
+ route_layer.add(route)
+ assert route_layer.index is not None and route_layer.categories is not None
+ assert len(route_layer.index) == 5
+ assert len(set(route_layer.categories)) == 2
+
+ def test_query_and_classification(self, openai_encoder, routes):
+ route_layer = HybridRouteLayer(encoder=openai_encoder, routes=routes)
+ query_result = route_layer("Hello")
+ assert query_result in ["Route 1", "Route 2"]
+
+ def test_query_with_no_index(self, openai_encoder):
+ route_layer = HybridRouteLayer(encoder=openai_encoder)
+ assert route_layer("Anything") is None
+
+ def test_semantic_classify(self, openai_encoder, routes):
+ route_layer = HybridRouteLayer(encoder=openai_encoder, routes=routes)
+ classification, score = route_layer._semantic_classify(
+ [
+ {"route": "Route 1", "score": 0.9},
+ {"route": "Route 2", "score": 0.1},
+ ]
+ )
+ assert classification == "Route 1"
+ assert score == [0.9]
+
+ def test_semantic_classify_multiple_routes(self, openai_encoder, routes):
+ route_layer = HybridRouteLayer(encoder=openai_encoder, routes=routes)
+ classification, score = route_layer._semantic_classify(
+ [
+ {"route": "Route 1", "score": 0.9},
+ {"route": "Route 2", "score": 0.1},
+ {"route": "Route 1", "score": 0.8},
+ ]
+ )
+ assert classification == "Route 1"
+ assert score == [0.9, 0.8]
+
+ def test_pass_threshold(self, openai_encoder):
+ route_layer = HybridRouteLayer(encoder=openai_encoder)
+ assert not route_layer._pass_threshold([], 0.5)
+ assert route_layer._pass_threshold([0.6, 0.7], 0.5)
+
+ def test_failover_score_threshold(self, base_encoder):
+ route_layer = HybridRouteLayer(encoder=base_encoder)
+ assert route_layer.score_threshold == 0.82
+
+
+# Add more tests for edge cases and error handling as needed.
diff --git a/tests/unit/test_layer.py b/tests/unit/test_layer.py
index d5f698bee02065e7e5f703c8a19faec67dbdf6fc..66e0d53bb9350c77578682f9ea0742b1d3dfe0b2 100644
--- a/tests/unit/test_layer.py
+++ b/tests/unit/test_layer.py
@@ -1,12 +1,7 @@
import pytest
from semantic_router.encoders import BaseEncoder, CohereEncoder, OpenAIEncoder
-from semantic_router.layer import (
- HybridRouteLayer,
- RouteLayer,
-)
-
-# Replace with the actual module name
+from semantic_router.layer import RouteLayer
from semantic_router.schema import Route
@@ -130,77 +125,4 @@ class TestRouteLayer:
assert route_layer.score_threshold == 0.82
-class TestHybridRouteLayer:
- def test_initialization(self, openai_encoder, routes):
- route_layer = HybridRouteLayer(encoder=openai_encoder, routes=routes)
- assert route_layer.index is not None and route_layer.categories is not None
- assert route_layer.score_threshold == 0.82
- assert len(route_layer.index) == 5
- assert len(set(route_layer.categories)) == 2
-
- def test_initialization_different_encoders(self, cohere_encoder, openai_encoder):
- route_layer_cohere = HybridRouteLayer(encoder=cohere_encoder)
- assert route_layer_cohere.score_threshold == 0.3
-
- route_layer_openai = HybridRouteLayer(encoder=openai_encoder)
- assert route_layer_openai.score_threshold == 0.82
-
- def test_add_route(self, openai_encoder):
- route_layer = HybridRouteLayer(encoder=openai_encoder)
- route = Route(name="Route 3", utterances=["Yes", "No"])
- route_layer.add(route)
- assert route_layer.index is not None and route_layer.categories is not None
- assert len(route_layer.index) == 2
- assert len(set(route_layer.categories)) == 1
-
- def test_add_multiple_routes(self, openai_encoder, routes):
- route_layer = HybridRouteLayer(encoder=openai_encoder)
- for route in routes:
- route_layer.add(route)
- assert route_layer.index is not None and route_layer.categories is not None
- assert len(route_layer.index) == 5
- assert len(set(route_layer.categories)) == 2
-
- def test_query_and_classification(self, openai_encoder, routes):
- route_layer = HybridRouteLayer(encoder=openai_encoder, routes=routes)
- query_result = route_layer("Hello")
- assert query_result in ["Route 1", "Route 2"]
-
- def test_query_with_no_index(self, openai_encoder):
- route_layer = HybridRouteLayer(encoder=openai_encoder)
- assert route_layer("Anything") is None
-
- def test_semantic_classify(self, openai_encoder, routes):
- route_layer = HybridRouteLayer(encoder=openai_encoder, routes=routes)
- classification, score = route_layer._semantic_classify(
- [
- {"route": "Route 1", "score": 0.9},
- {"route": "Route 2", "score": 0.1},
- ]
- )
- assert classification == "Route 1"
- assert score == [0.9]
-
- def test_semantic_classify_multiple_routes(self, openai_encoder, routes):
- route_layer = HybridRouteLayer(encoder=openai_encoder, routes=routes)
- classification, score = route_layer._semantic_classify(
- [
- {"route": "Route 1", "score": 0.9},
- {"route": "Route 2", "score": 0.1},
- {"route": "Route 1", "score": 0.8},
- ]
- )
- assert classification == "Route 1"
- assert score == [0.9, 0.8]
-
- def test_pass_threshold(self, openai_encoder):
- route_layer = HybridRouteLayer(encoder=openai_encoder)
- assert not route_layer._pass_threshold([], 0.5)
- assert route_layer._pass_threshold([0.6, 0.7], 0.5)
-
- def test_failover_score_threshold(self, base_encoder):
- route_layer = HybridRouteLayer(encoder=base_encoder)
- assert route_layer.score_threshold == 0.82
-
-
# Add more tests for edge cases and error handling as needed.
diff --git a/tests/unit/test_schema.py b/tests/unit/test_schema.py
index 7c954805bd0112fa7ce1e015aee716621faced12..f471755c35796d33ac9329a2ddb3a20816230cda 100644
--- a/tests/unit/test_schema.py
+++ b/tests/unit/test_schema.py
@@ -2,10 +2,10 @@ import pytest
from semantic_router.schema import (
CohereEncoder,
- Route,
Encoder,
EncoderType,
OpenAIEncoder,
+ Route,
SemanticSpace,
)
diff --git a/walkthrough.ipynb b/walkthrough.ipynb
index a4265e5a1dfc9b2dd5060d54b54add8e4cc67e6d..d31a88dc0da38dec04e019100bfc28815adf7906 100644
--- a/walkthrough.ipynb
+++ b/walkthrough.ipynb
@@ -128,7 +128,7 @@
"metadata": {},
"outputs": [],
"source": [
- "from semantic_router.layer import RouteLayer\n",
+ "from semantic_router.router import RouteLayer\n",
"\n",
"dl = RouteLayer(encoder=encoder, routes=routes)"
]