feat: toggle hybrid search

This commit is contained in:
Steven Kreitzer 2024-04-25 17:31:21 -05:00
parent 984dbf13ab
commit 9755cd5baa
4 changed files with 133 additions and 88 deletions

View file

@ -70,6 +70,7 @@ from config import (
RAG_EMBEDDING_MODEL,
RAG_EMBEDDING_MODEL_AUTO_UPDATE,
RAG_EMBEDDING_MODEL_TRUST_REMOTE_CODE,
RAG_HYBRID,
RAG_RERANKING_MODEL,
RAG_RERANKING_MODEL_AUTO_UPDATE,
RAG_RERANKING_MODEL_TRUST_REMOTE_CODE,
@ -91,6 +92,8 @@ app = FastAPI()
app.state.TOP_K = RAG_TOP_K
app.state.RELEVANCE_THRESHOLD = RAG_RELEVANCE_THRESHOLD
app.state.HYBRID = RAG_HYBRID
app.state.CHUNK_SIZE = CHUNK_SIZE
app.state.CHUNK_OVERLAP = CHUNK_OVERLAP
@ -321,6 +324,7 @@ async def get_query_settings(user=Depends(get_admin_user)):
"template": app.state.RAG_TEMPLATE,
"k": app.state.TOP_K,
"r": app.state.RELEVANCE_THRESHOLD,
"hybrid": app.state.HYBRID,
}
@ -328,6 +332,7 @@ class QuerySettingsForm(BaseModel):
k: Optional[int] = None
r: Optional[float] = None
template: Optional[str] = None
hybrid: Optional[bool] = None
@app.post("/query/settings/update")
@ -337,7 +342,14 @@ async def update_query_settings(
app.state.RAG_TEMPLATE = form_data.template if form_data.template else RAG_TEMPLATE
app.state.TOP_K = form_data.k if form_data.k else 4
app.state.RELEVANCE_THRESHOLD = form_data.r if form_data.r else 0.0
return {"status": True, "template": app.state.RAG_TEMPLATE}
app.state.HYBRID = form_data.hybrid if form_data.hybrid else False
return {
"status": True,
"template": app.state.RAG_TEMPLATE,
"k": app.state.TOP_K,
"r": app.state.RELEVANCE_THRESHOLD,
"hybrid": app.state.HYBRID,
}
class QueryDocForm(BaseModel):
@ -345,6 +357,7 @@ class QueryDocForm(BaseModel):
query: str
k: Optional[int] = None
r: Optional[float] = None
hybrid: Optional[bool] = None
@app.post("/query/doc")
@ -368,6 +381,7 @@ def query_doc_handler(
r=form_data.r if form_data.r else app.state.RELEVANCE_THRESHOLD,
embeddings_function=embeddings_function,
reranking_function=app.state.sentence_transformer_rf,
hybrid=form_data.hybrid if form_data.hybrid else app.state.HYBRID,
)
except Exception as e:
log.exception(e)
@ -382,6 +396,7 @@ class QueryCollectionsForm(BaseModel):
query: str
k: Optional[int] = None
r: Optional[float] = None
hybrid: Optional[bool] = None
@app.post("/query/collection")
@ -405,6 +420,7 @@ def query_collection_handler(
r=form_data.r if form_data.r else app.state.RELEVANCE_THRESHOLD,
embeddings_function=embeddings_function,
reranking_function=app.state.sentence_transformer_rf,
hybrid=form_data.hybrid if form_data.hybrid else app.state.HYBRID,
)
except Exception as e:
log.exception(e)

View file

@ -32,13 +32,13 @@ def query_embeddings_doc(
collection_name: str,
query: str,
embeddings_function,
reranking_function,
k: int,
reranking_function: Optional[CrossEncoder] = None,
r: Optional[float] = None,
hybrid: Optional[bool] = False,
):
try:
if reranking_function:
if hybrid:
# if you use docker use the model from the environment variable
collection = CHROMA_CLIENT.get_collection(name=collection_name)
@ -142,6 +142,7 @@ def query_embeddings_collection(
r: float,
embeddings_function,
reranking_function,
hybrid: bool,
):
results = []
@ -155,6 +156,7 @@ def query_embeddings_collection(
r=r,
embeddings_function=embeddings_function,
reranking_function=reranking_function,
hybrid=hybrid,
)
results.append(result)
except:
@ -211,6 +213,7 @@ def rag_messages(
template,
k,
r,
hybrid,
embedding_engine,
embedding_model,
embedding_function,
@ -283,6 +286,7 @@ def rag_messages(
r=r,
embeddings_function=embeddings_function,
reranking_function=reranking_function,
hybrid=hybrid,
)
else:
context = query_embeddings_doc(
@ -292,6 +296,7 @@ def rag_messages(
r=r,
embeddings_function=embeddings_function,
reranking_function=reranking_function,
hybrid=hybrid,
)
except Exception as e:
log.exception(e)

View file

@ -422,6 +422,7 @@ CHROMA_DATA_PATH = f"{DATA_DIR}/vector_db"
RAG_TOP_K = int(os.environ.get("RAG_TOP_K", "5"))
RAG_RELEVANCE_THRESHOLD = float(os.environ.get("RAG_RELEVANCE_THRESHOLD", "0.0"))
RAG_HYBRID = os.environ.get("RAG_HYBRID", "").lower() == "true"
RAG_EMBEDDING_ENGINE = os.environ.get("RAG_EMBEDDING_ENGINE", "")

View file

@ -43,7 +43,8 @@
let querySettings = {
template: '',
r: 0.0,
k: 4
k: 4,
hybrid: false
};
const scanHandler = async () => {
@ -174,6 +175,12 @@
}
};
const toggleHybridSearch = async () => {
querySettings.hybrid = !querySettings.hybrid;
querySettings = await updateQuerySettings(localStorage.token, querySettings);
};
onMount(async () => {
const res = await getRAGConfig(localStorage.token);
@ -202,6 +209,24 @@
<div>
<div class=" mb-2 text-sm font-medium">{$i18n.t('General Settings')}</div>
<div class=" flex w-full justify-between">
<div class=" self-center text-xs font-medium">{$i18n.t('Hybrid Search')}</div>
<button
class="p-1 px-3 text-xs flex rounded transition"
on:click={() => {
toggleHybridSearch();
}}
type="button"
>
{#if querySettings.hybrid === true}
<span class="ml-2 self-center">{$i18n.t('On')}</span>
{:else}
<span class="ml-2 self-center">{$i18n.t('Off')}</span>
{/if}
</button>
</div>
<div class=" flex w-full justify-between">
<div class=" self-center text-xs font-medium">{$i18n.t('Embedding Model Engine')}</div>
<div class="flex items-center relative">
@ -386,78 +411,74 @@
<hr class=" dark:border-gray-700 my-3" />
<div class=" ">
<div class=" mb-2 text-sm font-medium">{$i18n.t('Update Reranking Model')}</div>
{#if querySettings.hybrid === true}
<div class=" ">
<div class=" mb-2 text-sm font-medium">{$i18n.t('Update Reranking Model')}</div>
<div class="flex w-full">
<div class="flex-1 mr-2">
<input
class="w-full rounded-lg py-2 px-4 text-sm dark:text-gray-300 dark:bg-gray-850 outline-none"
placeholder={$i18n.t('Update reranking model (e.g. {{model}})', {
model: rerankingModel.slice(-40)
})}
bind:value={rerankingModel}
/>
</div>
<button
class="px-2.5 bg-gray-100 hover:bg-gray-200 text-gray-800 dark:bg-gray-850 dark:hover:bg-gray-800 dark:text-gray-100 rounded-lg transition"
on:click={() => {
rerankingModelUpdateHandler();
}}
disabled={updateRerankingModelLoading}
>
{#if updateRerankingModelLoading}
<div class="self-center">
<svg
class=" w-4 h-4"
viewBox="0 0 24 24"
fill="currentColor"
xmlns="http://www.w3.org/2000/svg"
><style>
.spinner_ajPY {
transform-origin: center;
animation: spinner_AtaB 0.75s infinite linear;
}
@keyframes spinner_AtaB {
100% {
transform: rotate(360deg);
<div class="flex w-full">
<div class="flex-1 mr-2">
<input
class="w-full rounded-lg py-2 px-4 text-sm dark:text-gray-300 dark:bg-gray-850 outline-none"
placeholder={$i18n.t('Update reranking model (e.g. {{model}})', {
model: rerankingModel.slice(-40)
})}
bind:value={rerankingModel}
/>
</div>
<button
class="px-2.5 bg-gray-100 hover:bg-gray-200 text-gray-800 dark:bg-gray-850 dark:hover:bg-gray-800 dark:text-gray-100 rounded-lg transition"
on:click={() => {
rerankingModelUpdateHandler();
}}
disabled={updateRerankingModelLoading}
>
{#if updateRerankingModelLoading}
<div class="self-center">
<svg
class=" w-4 h-4"
viewBox="0 0 24 24"
fill="currentColor"
xmlns="http://www.w3.org/2000/svg"
><style>
.spinner_ajPY {
transform-origin: center;
animation: spinner_AtaB 0.75s infinite linear;
}
}
</style><path
d="M12,1A11,11,0,1,0,23,12,11,11,0,0,0,12,1Zm0,19a8,8,0,1,1,8-8A8,8,0,0,1,12,20Z"
opacity=".25"
/><path
d="M10.14,1.16a11,11,0,0,0-9,8.92A1.59,1.59,0,0,0,2.46,12,1.52,1.52,0,0,0,4.11,10.7a8,8,0,0,1,6.66-6.61A1.42,1.42,0,0,0,12,2.69h0A1.57,1.57,0,0,0,10.14,1.16Z"
class="spinner_ajPY"
/></svg
@keyframes spinner_AtaB {
100% {
transform: rotate(360deg);
}
}
</style><path
d="M12,1A11,11,0,1,0,23,12,11,11,0,0,0,12,1Zm0,19a8,8,0,1,1,8-8A8,8,0,0,1,12,20Z"
opacity=".25"
/><path
d="M10.14,1.16a11,11,0,0,0-9,8.92A1.59,1.59,0,0,0,2.46,12,1.52,1.52,0,0,0,4.11,10.7a8,8,0,0,1,6.66-6.61A1.42,1.42,0,0,0,12,2.69h0A1.57,1.57,0,0,0,10.14,1.16Z"
class="spinner_ajPY"
/></svg
>
</div>
{:else}
<svg
xmlns="http://www.w3.org/2000/svg"
viewBox="0 0 16 16"
fill="currentColor"
class="w-4 h-4"
>
</div>
{:else}
<svg
xmlns="http://www.w3.org/2000/svg"
viewBox="0 0 16 16"
fill="currentColor"
class="w-4 h-4"
>
<path
d="M8.75 2.75a.75.75 0 0 0-1.5 0v5.69L5.03 6.22a.75.75 0 0 0-1.06 1.06l3.5 3.5a.75.75 0 0 0 1.06 0l3.5-3.5a.75.75 0 0 0-1.06-1.06L8.75 8.44V2.75Z"
/>
<path
d="M3.5 9.75a.75.75 0 0 0-1.5 0v1.5A2.75 2.75 0 0 0 4.75 14h6.5A2.75 2.75 0 0 0 14 11.25v-1.5a.75.75 0 0 0-1.5 0v1.5c0 .69-.56 1.25-1.25 1.25h-6.5c-.69 0-1.25-.56-1.25-1.25v-1.5Z"
/>
</svg>
{/if}
</button>
<path
d="M8.75 2.75a.75.75 0 0 0-1.5 0v5.69L5.03 6.22a.75.75 0 0 0-1.06 1.06l3.5 3.5a.75.75 0 0 0 1.06 0l3.5-3.5a.75.75 0 0 0-1.06-1.06L8.75 8.44V2.75Z"
/>
<path
d="M3.5 9.75a.75.75 0 0 0-1.5 0v1.5A2.75 2.75 0 0 0 4.75 14h6.5A2.75 2.75 0 0 0 14 11.25v-1.5a.75.75 0 0 0-1.5 0v1.5c0 .69-.56 1.25-1.25 1.25h-6.5c-.69 0-1.25-.56-1.25-1.25v-1.5Z"
/>
</svg>
{/if}
</button>
</div>
</div>
</div>
<div class="mt-2 mb-1 text-xs text-gray-400 dark:text-gray-500">
{$i18n.t(
'Note: If you choose a reranking model, it will use that to score and rerank instead of the embedding model.'
)}
</div>
<hr class=" dark:border-gray-700 my-3" />
<hr class=" dark:border-gray-700 my-3" />
{/if}
<div class=" flex w-full justify-between">
<div class=" self-center text-xs font-medium">
@ -583,25 +604,27 @@
</div>
</div>
<div class=" flex">
<div class=" flex w-full justify-between">
<div class="self-center text-xs font-medium flex-1">
{$i18n.t('Relevance Threshold')}
</div>
{#if querySettings.hybrid === true}
<div class=" flex">
<div class=" flex w-full justify-between">
<div class="self-center text-xs font-medium flex-1">
{$i18n.t('Relevance Threshold')}
</div>
<div class="self-center p-3">
<input
class=" w-full rounded-lg py-1.5 px-4 text-sm dark:text-gray-300 dark:bg-gray-850 outline-none"
type="number"
step="0.01"
placeholder={$i18n.t('Enter Relevance Threshold')}
bind:value={querySettings.r}
autocomplete="off"
min="0.0"
/>
<div class="self-center p-3">
<input
class=" w-full rounded-lg py-1.5 px-4 text-sm dark:text-gray-300 dark:bg-gray-850 outline-none"
type="number"
step="0.01"
placeholder={$i18n.t('Enter Relevance Threshold')}
bind:value={querySettings.r}
autocomplete="off"
min="0.0"
/>
</div>
</div>
</div>
</div>
{/if}
<div>
<div class=" mb-2.5 text-sm font-medium">{$i18n.t('RAG Template')}</div>