From 8d221e4d82f03fc0f24ca8705a5b70438117ca3c Mon Sep 17 00:00:00 2001
From: Surya <git@sdan.io>
Date: Fri, 12 Apr 2024 16:24:59 -0700
Subject: [PATCH 1/2] add sentence-transformers uses MRL

---
 requirements.txt |  3 +-
 vlite/main.py    | 30 ++++++++--------
 vlite/model.py   | 90 +++++++++++++++++++++---------------------------
 3 files changed, 56 insertions(+), 67 deletions(-)

diff --git a/requirements.txt b/requirements.txt
index 8226f39..602ca71 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -8,4 +8,5 @@ llama-cpp-python
 huggingface_hub
 tiktoken
 onnxruntime==1.17.1
-tokenizers==0.15.2
\ No newline at end of file
+tokenizers==0.15.2
+sentence_transformers
\ No newline at end of file
diff --git a/vlite/main.py b/vlite/main.py
index 6d3c456..70c8b60 100644
--- a/vlite/main.py
+++ b/vlite/main.py
@@ -86,16 +86,19 @@ def retrieve(self, text=None, top_k=5, metadata=None, return_scores=False):
         print("Retrieving similar texts...")
         if text:
             print(f"Retrieving top {top_k} similar texts for query: {text}")
-            query_chunks = chop_and_chunk(text, fast=True)
-            query_vectors = self.model.embed(query_chunks, device=self.device)
+            
+            # Embed and quantize the query text
+            query_vectors = self.model.embed(text, device=self.device)
             query_binary_vectors = self.model.quantize(query_vectors, precision="binary")
             
+            # Perform search on the query binary vectors
             results = []
             for query_binary_vector in query_binary_vectors:
-                chunk_results = self.search(query_binary_vector, top_k, metadata)
+                chunk_results = self.rank_and_filter(query_binary_vector, top_k, metadata)
                 results.extend(chunk_results)
             
-            results.sort(key=lambda x: x[1], reverse=True)
+            # Sort the results by similarity score 
+            results.sort(key=lambda x: x[1])
             results = results[:top_k]
             
             print("Retrieval completed.")
@@ -103,15 +106,12 @@ def retrieve(self, text=None, top_k=5, metadata=None, return_scores=False):
                 return [(idx, self.index[idx]['text'], self.index[idx]['metadata'], score) for idx, score in results]
             else:
                 return [(idx, self.index[idx]['text'], self.index[idx]['metadata']) for idx, _ in results]
-
-    def search(self, query_binary_vector, top_k, metadata=None):
-        # Reshape query_binary_vector to 1D array
-        query_binary_vector = query_binary_vector.reshape(-1)
-
-        # Perform binary search
-        binary_vectors = np.array([item['binary_vector'] for item in self.index.values()])
-        binary_similarities = np.einsum('i,ji->j', query_binary_vector, binary_vectors)
-        top_k_indices = np.argpartition(binary_similarities, -top_k)[-top_k:]
+        
+    def rank_and_filter(self, query_binary_vector, top_k, metadata=None):
+        query_binary_vector = np.array(query_binary_vector).reshape(-1)
+        
+        corpus_binary_vectors = np.array([item['binary_vector'] for item in self.index.values()])
+        top_k_indices, top_k_scores = self.model.search(query_binary_vector, corpus_binary_vectors, top_k)
         top_k_ids = [list(self.index.keys())[idx] for idx in top_k_indices]
 
         # Apply metadata filter on the retrieved top_k items
@@ -122,9 +122,7 @@ def search(self, query_binary_vector, top_k, metadata=None):
                 if all(item_metadata.get(key) == value for key, value in metadata.items()):
                     filtered_ids.append(chunk_id)
             top_k_ids = filtered_ids[:top_k]
-
-        # Get the similarity scores for the top_k items
-        top_k_scores = binary_similarities[top_k_indices]
+            top_k_scores = top_k_scores[:len(top_k_ids)]
 
         return list(zip(top_k_ids, top_k_scores))
 
diff --git a/vlite/model.py b/vlite/model.py
index 054ea13..64970e9 100644
--- a/vlite/model.py
+++ b/vlite/model.py
@@ -4,71 +4,61 @@
 import numpy as np
 from typing import List
 from tokenizers import Tokenizer
+import numpy as np
+from typing import List
+from sentence_transformers import SentenceTransformer
+
 
 def normalize(v):
-    norm = np.linalg.norm(v, axis=1)
+    if v.ndim == 1:
+        v = v.reshape(1, -1)  # Reshape v to 2D array if it is 1D
+    norm = np.linalg.norm(v, axis=1, keepdims=True)
     norm[norm == 0] = 1e-12
-    return v / norm[:, np.newaxis]
+    return v / norm
+
+
 
 class EmbeddingModel:
     def __init__(self, model_name="mixedbread-ai/mxbai-embed-large-v1"):
-        tokenizer_path = hf_hub_download(repo_id=model_name, filename="tokenizer.json")
-        model_path = hf_hub_download(repo_id=model_name, filename="onnx/model.onnx")
-        
-        self.tokenizer = Tokenizer.from_file(tokenizer_path)
-        self.tokenizer.enable_truncation(max_length=512)
-        self.tokenizer.enable_padding(pad_id=0, pad_token="[PAD]", length=512)
-        
-        self.model = ort.InferenceSession(model_path)
-        print("[model]", self.model.get_modelmeta())
-        
+        self.model = SentenceTransformer(model_name)
         self.model_metadata = {
-            "bert.embedding_length": 1024,
+            "bert.embedding_length": 512,
             "bert.context_length": 512
         }
         self.embedding_size = self.model_metadata.get("bert.embedding_length", 1024)
         self.context_length = self.model_metadata.get("bert.context_length", 512)
         self.embedding_dtype = "float32"
+    
+    def embed(self, texts, max_seq_length=512, device="cpu", batch_size=32):
+        if isinstance(texts, str):
+            texts = [texts]  # Ensure texts is always a list
+        embeddings = self.model.encode(texts, device=device, batch_size=batch_size, normalize_embeddings=True)
+        return embeddings
 
-    def embed(self, texts: List[str], max_seq_length=512, device="cpu", batch_size=32):
-        all_embeddings = []
-        for i in range(0, len(texts), batch_size):
-            batch = texts[i:i + batch_size]
-            encoded = [self.tokenizer.encode(d) for d in batch]
-            input_ids = np.array([e.ids for e in encoded])
-            attention_mask = np.array([e.attention_mask for e in encoded])
-            token_type_ids = np.zeros_like(input_ids, dtype=np.int64)
-
-            onnx_input = {
-                "input_ids": input_ids,
-                "attention_mask": attention_mask,
-                "token_type_ids": token_type_ids,
-            }
-            model_output = self.model.run(None, onnx_input)
-            last_hidden_state = model_output[0]
-
-            input_mask_expanded = np.broadcast_to(np.expand_dims(attention_mask, -1), last_hidden_state.shape)
-            embeddings = np.sum(last_hidden_state * input_mask_expanded, 1) / np.clip(input_mask_expanded.sum(1), a_min=1e-9, a_max=None)
-            embeddings = normalize(embeddings).astype(np.float32)
-            all_embeddings.append(embeddings)
-
-        return np.concatenate(all_embeddings)
-
-    def token_count(self, texts):
-        tokens = 0
-        for text in texts:
-            encoded = self.tokenizer.encode(text)
-            tokens += len(encoded.ids)
-        return tokens
 
     def quantize(self, embeddings, precision="binary"):
-        embeddings = np.array(embeddings)
+        # first normalize_embeddings to unit length
+        embeddings = normalize(embeddings)
+        # slice to get MRL embeddings
+        embeddings_slice = embeddings[..., :512]
+        
         if precision == "binary":
-            return np.packbits(embeddings > 0).reshape(embeddings.shape[0], -1)
-        elif precision == "int8":
-            return ((embeddings - np.min(embeddings, axis=0)) / (np.max(embeddings, axis=0) - np.min(embeddings, axis=0)) * 255).astype(np.uint8)
+            return self._binary_quantize(embeddings_slice)
         else:
-            raise ValueError(f"Unsupported precision: {precision}")
+            raise ValueError(f"Precision {precision} is not supported")
 
-    def rescore(self, query_vector, vectors):
-        return np.dot(query_vector, vectors.T).flatten()
\ No newline at end of file
+    def _binary_quantize(self, embeddings):
+        return (np.packbits(embeddings > 0).reshape(embeddings.shape[0], -1) - 128).astype(np.int8)
+
+    def hamming_distance(self, embedding1, embedding2):
+        # Ensure the embeddings are numpy arrays for the operation.
+        return np.count_nonzero(np.array(embedding1) != np.array(embedding2))
+
+    def search(self, query_embedding, embeddings, top_k):
+        # Convert embeddings to a numpy array for efficient operations if not already.
+        embeddings = np.array(embeddings)
+        distances = np.array([self.hamming_distance(query_embedding, emb) for emb in embeddings])
+        
+        # Find the indices of the top_k smallest distances
+        top_k_indices = np.argsort(distances)[:top_k]
+        return top_k_indices, distances[top_k_indices]
\ No newline at end of file

From 8cc1ed2c659c34698fac87eb782e94d3d95ddbb5 Mon Sep 17 00:00:00 2001
From: Surya <git@sdan.io>
Date: Fri, 12 Apr 2024 17:10:37 -0700
Subject: [PATCH 2/2] implementing onnx

---
 requirements.txt                 |   4 +-
 tests/contexts/my_collection.ctx | Bin 0 -> 48604 bytes
 vlite/main.py                    |   4 +-
 vlite/model.py                   |  76 ++++++++++++++++++++++++++-----
 4 files changed, 67 insertions(+), 17 deletions(-)
 create mode 100644 tests/contexts/my_collection.ctx

diff --git a/requirements.txt b/requirements.txt
index 602ca71..820f4c9 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -4,9 +4,7 @@ docx2txt
 pandas
 Requests
 beautifulsoup4
-llama-cpp-python
 huggingface_hub
 tiktoken
 onnxruntime==1.17.1
-tokenizers==0.15.2
-sentence_transformers
\ No newline at end of file
+tokenizers==0.15.2
\ No newline at end of file
diff --git a/tests/contexts/my_collection.ctx b/tests/contexts/my_collection.ctx
new file mode 100644
index 0000000000000000000000000000000000000000..a58a7fcf123a469e9741767683f3c8750282bfc0
GIT binary patch
literal 48604
zcmb8Y3y5Ubd8pfxl~}1b@g$K?<XCq0QIzAJxVzoYnVy!$>3K;SX<nWgY4ki+byrPy
zxx1^{^_ZC+OL0e-91(nj@HIg<h{Iq^t_a46V1h9s46$zz!59%PK?D(mF~J0rBZ9dY
zBM$fb{<Zh6s-BjeY384`pKJZ=f3JV7z3RyM$L_ynB>nvf#pmy?Z!FbowMKL0GaId1
zef{o*?Cy=mm3nQdU9Z;0tBp@?Tv@6%#^bT^^=f;iK7MKH?y+p(#ZKdDo%g0Dr)P(*
z*ScGq-njeXdaK%<nf9&aR<m2b(*4YOy}8m|<>uVf^mBpHyjVSSWMt%>!y_Zxe>gJo
zF6Hf<-%0sjDF1tYZ$IT(;9G}AMt;V<5#SEaYm{#Se+T?2@aMq!{JGzw{6pYtT)RN|
zb;`GZ%N*Ozv%u@rSLgf;Fv;<cfaU!7pSf@QFHydq@=0I?_&D(IT)T(zlfZMp%bXve
z{7vBBIG&*V9pIlh9;N&trS%=;_<71Rz<&VO7Dq;YK=~@|j#7%d*QLoe<p(TzW{PqI
z5dSAA@1nd7_%hJoTsnyRgB*X4<2}G1aeN=}O^!c8`32xB9B%^GsMBk|0{l8)pU!jr
zqm&Ck1^5i`zquyerR{H1+V5K_=P6$S7OCfH&QAb)0Q>4QUCO(GL%=_B?z{Fa1IKAf
z`*59i#qSib3cSd3e+!K8%<Yu>xF#JQ18xBx0c`81+_xY0_j8<YQvM6)e+8_0pYjuw
z^7$yo8Ra)9-v+)9oa6jC;9lTcoIgM*?$YVQlrz9}a1$4KE<JyV^7DXn?@=mmvy{^9
zZQ8dFUcZg&{|ZQlj{(*#&n{8EN%>zX4+Bep{goc#^C4iHHjh!tpAULYsT}#YWy$#@
za3{x)Q7T8@25du}xK3U6^@|)^TK_$OGIWj7K8vT%-NE^9ar`*20Z6Z>xMrVi>(jvd
zfF|%mu8Eg8ewX8s#e~a?)U(L3xGHzHAupuYan9|l?fea3|9=Bej(pZ@>f?$0{3n#R
zb6r{~d;bl%PTlI?8(g!$h5oBY%7^kKu0Q4YtAKT@5As8K>i~1W*SIGCzX+(0+Jyg2
z`7E%_vlltP0!(r&J+0?Wj&0{P;03^bDGMvWZJb-TG?xF;=4GxOpwwpBXKlk9ydxd{
z0B8g9-)FWd#aTV~?+28x@@z(FpS-TzzfGwOUE^NpKj&9DS3mFKcm#Ny<3j&^PTGE#
z>&l7pdJy;>Kwh_iZ*qMTxVD(&Uz-0ipskcv>eQzw#aSHQ<=H{~_gwwIkMbzxBK4?~
z_HzWV59+@(QU9e?q5s;8XQ=;&l=kB~@2dZ{Bmb3aX{5aRt~8PV>c2Q2rWC(||6R_d
z=N|5<bC!PwJjJySP)f6FwDC2LwFP%lioY~b7pznMuX9}3|G(zi66ce7JGT8LuDze~
z&ncy$GSCC$vB&DlSGc!DDGj`L8}KUE9;RFc9^_cOZTriV|1Th){o6xZF9OnTk^AD{
zd)t)blrM14ekyO)B|q=t-W`DQs*Qby@<WvJ-+p@Uk2pS0DGlYndj1h0`i-xh;rLg8
zUkBv>en4K%10#zgBd00l(@y>?FW;kH`?$z+Cn@zUBY7X>zp|nIr~=Bmbv)uR&nT}?
z0P<H^Tcmu4X9oSh`tSXZQF`ww;ERC%-#(83)^`P%<oNG_Yv7{XsY5;|U+mM@xhMaX
zzqbJOU)uZ|$I|V7jyD1Mt9*&`B6Z2HZvpo8x4EXxlK;|5{htNYfAR1+d8GbZud*SZ
zq;JOg-IU@vO8GqH?STCMCC)zw*w<0c<-ga}iET>lzjRTSt@8@k3jW{5v9cmPl|6Bn
z|KH`>9zfoyUoUff3#I-33dftkHTrii=X-cZ{@Xug<)f6=Ujd|xyi<R)i9Yi#r7}KG
zc?abP^~is9P9N%f^1nxEpI_uyI-cWLzF7A{UfO5zyB!GK;oSatz63~*Q<U=f8u!FQ
z`O$XEUvd5aINk<_f1<Q+@?X5}<5>GAf8_fGO8HpWef!Y`q(#C1P3n|}f5$cd)E9N<
zPR<Mdd+jdH3;zE;=Uw0(-jk2=U!IIn+Rx7b_TzToO<<h!>y&~r_yVQ86%XY^{wrtF
z)O+$@{hy?i?$#j<#YsHH=Q!{L@Cwh#|1WX<8II+{Bw(LD#C7XZu59ZLU>#6C<-f8q
z2RsK{1v21+fb{wyQ1Jg*uDwWo^IQ`@`5><1aVKCKz9X-vfklqxzqa5N%3r3`1}S?5
z|HWCEKSN#iPyWjj`FIMj%`;rrhRA>GyB)AkHA?S|0nY<}4%oJ|(7rvzxpE^dz6ksa
zAm8ohvy{?8da3`Dyz_02rK`052TJ+BhthsN$+7w`|HWk%P`(}oUgMg4^w|34zkFH)
zMmbj(UZ>PX$ouaC%C&#;SGn;yako$M=@m-r@xIT>lR^G(FDAOH|I$x=kd{B@*n4e`
zCprEY<!=Gv{!Py9^EK|vb7?LA?Z<hpKTY{5z_#8G$V=&0@L#&!3cSTNZTuqj_|D^i
z{I@=7Z9nYuA|Stf&i+Uf`?dfS{NLvK+kpIknNl0#wcn(a{?>Vh^Cu|Z<e8sw{6j$g
zOBelz{QoyV{)&%%Q%98n>8}3UZ|@)E*nT|={0Mk}bK93T>dKPm;4A-M1@-{iX>HUv
zumVg21^=yEyng_Atp0z1@(t>`jq(q2kVn$>?<k)KJhyMYC;zYWyfhV0X>^TJ-e~`&
zm-av7+B~KDFFlo4b<euQ-TIWR5%B*k$Li?`N@ZKTt@j;{@8G(6Xq(>Gr|kpOHT&SZ
z>gu~ZC#?(qU*g&bbxX&u126HuHcr~e|52{_jx>;eZ}Y7Df0gI#qwW7C=i;N@$&2fh
z?*M1G_7>#`<wdUF0sIm0qW6Gv9RDrlbClxr2&L`r;rJdv9OZ+u_Y|eP-w%lAHvrqw
zUljaT2CVnbfOmjBv?~pldB;ETehqj5_-(Fz9=HrhdvTHf(oH;G;@Sa9+vx)GUs>0d
z{{f}C@FvIlJ9+Q%;~Z~ON-OCjZ_jY8Mk$@HQ5O7{=gQ6)Af3g*cjdqOuRRs_2Pvh^
z^~KcYB**H|ZvpDRI<(5QYupz<>k?P{A+4mXvMtWayfjiT3jTkAYvS>JuF0Q8>Uo+{
zyo3Lg|IWQvff4Tc_m3QFOYGx2z}Glep06(^UhLuAXXN`&DV1UAWhw6R#=iNU^!zK}
zK|oyfxzBJ<zWxZ1XLDSerIi2bw({_4;90IIKh`DvuW;>Kl=_+%Db<CKaxCuK;AUM<
zasE9@@n5G@{?(Hsz{de;{|vCnyFE&2_EnyfruNn6?&18qfVhux{$<K90zU(kP5FNp
z*GDK%Q2s0BuTje12}|lxFV%nF3BFNEH_wg#|B3QLlm-8#-zqT4b^B&n=ln51dfmpk
z`04)-0|o!>uQDQ!y>}Jxe!+ijjqR#4Z$l&f;%TlaueP0WU71<rx-zGZ`~~M<1LT7;
zqx_BVu6^GGeD58~GrS{TZs%P6zs@tKDBs}PV}N!1lxI$IP5vu$>ZI?C0n*`rQreEZ
zev<MoO5al^l@D?Ap7cCVxd_-+;s2El<A48N=brNSJxcXo{uTbe(Eo9++uqCE)BkJZ
z)PH&PF+d%cCh}vFQvH>$_DPxb?+28=H%e(;HA?B}bMpTMaGIwa0qkqR|L=0F?kR)+
z!SO3RCoRQydvVzR8~<C^SGczcT;u)%=koYAKv`D*m4C0x{|^H1^6Xj4g8%COt-xDc
z7uQAV5eGs2uK?28d-nM{$Nvnxk7KX>I^Z+P)vMrpAEmT-8@NVYwx{mMlWorJ>mnep
zmMN8S>uXca0@{sv;O~Lo=UiId;XTgnk8q7^-{4%j$U|cT`ytPjIptSbn&jC2e-^k-
z+qR=V*jM>)>>>ZP1@=cfAl=lfx49?(|AA-azkPUvYrjgFaeg}>p8AMw%6kBLHo`MU
zDW%12lt;OLjj|a3%d>6nO;TRxy8Krrt#2C~Kg_XwQg&=t{{NU#{)ww?pQ03B>0|%p
z{~q93K)FyR9{}#+{2=8!v?q_PR~_<L{{JrJzfg|Qt~3$XMUK@+c|XX1>1Lba`z^p&
zK>kaU)x5ON@?V_a=9x2`ze{;L=gN-fpQ3z+Ys&VY<fS@n{~zJr7VtX9HQ*Y!sQ>cc
zm`VFD-?q6X|Gxx?w{^<@=K%Sp{Z|&`#kaYC9e5rX<iCBFb}N+fT)wK~p9JK;eU!If
zr5xnHetC>*e?Y067yOrB@?YG<Q(BC0EKPqvDbKv892NYRr$6P|U6ji5b^7}_$KtXE
zoadUf^qzHjP5P>H@<Lnn1n>;^uJNvVF0F3^3RzbE-vJl#l`h(-!<6!0doJCqOF0nN
zNuHJe&jRw__vDZK_PG~1RtAg{<%|0KEa#T?+4rQMe3Tc`QJnmfZsKwerLyyvlpRWO
z_zB0pTLZq&v3mD+fcD@FpkGq=<&XXI-5&uyD}S0?^S<rbzPuS(?AuUf%07o&Q~naq
zC}UoK45)BE2dHDu0GqTYZodrJSNV2@=bomN*LMPc4g6Qm<?HSF`PV6x8FkRV?L~OV
zx$l27FO_Bacn|O<*X+Ob{TpEY;-(MK-)Jl4|E-k1FMf{!%AK*&XMi7p-yZJCE9FPL
zd|tjBr2Hemcix{LTjvYh`(4T(QQC+5D8<XZMjyE*-wOV}$aUpq51<VE5Rm`M#Wav{
z|4Urk1nleEl<MERl-8rH%lo@5xhAcjr<AXS{#&2@dWq}G${m0*r>v->Gl2Rpjn)6x
zfwzD^<eKgDD6dmrq5t9}ZI-$I=iFESUjuy4dendEsk|6>$Pe2&28e^WNUQT4D|5>D
z2LW+g1*G|2&aYEufco#Be0i02y>^siuRlz=O({>Mvo!PlNuZEP`z!B`b5A|}5aolw
zcRBXis~jsg&rr8?`6TcJu*tn^i%I{jOZ#9yq_O<B&AWjW;3qtLm{MJk2GVVW@;=IK
z>KFmk|EB=!n&<kz1M<x}?CZ~fTRDD(auQrV!SP4D$GhUU3jBb3zW)}b^5DD5oA_P<
zx?HypKjnROSoxCATY!C)W?u!AZ?E6O@i|I;vUGfaQW{&ob!q3H<(mALcQuaBa9rc~
znf#qW`!63~=e{)bS!HV$`2GA`8cV;oDea3g{07I?A^+`<?a2S<DMtb6;x*}d2Qb31
z{Fj%fDX&pVpBJb{8iY=8O}V<0@)fSzk3*dM&KW@7{FqXj$cI@#nn*isp8h=Jnt%Ex
zpO@as*M8umfc1YCu+LWkahLxWD8B%VET&lD8tpG~{1@C4KkFI=lo9!)uFD7e_$PpT
z@%pb)ev@<Cd4^J%eTQ<9d%i2}Y*YHZ#<};@%fnog7V7^9?JDo;+?$-+jyS%b^QVCO
zIQ~6KpZOxk;{F)r%fKIUEbWEs)OXQ)zzD~-Q{`Cw`)BTZeUj38)iwFAUUUHY@oRwg
z!#;|OapX~srRkl3eo}h>1h5{Td7e_7{|S(0zXiNQyFMq4^i>a0Zc~m?+GpcG|9o$O
z<GUyYanc^wD8+q|<2yOt1Bj#Y^BTwUU!FbzY;vwmk^jE|h^xF;mbK;LtsPfZ#9O)-
zHc)xE6|imjKeCwg|2eM#@>Bl&6|ltlEr4{ief#z%@2LOsUir}8DC_^qGjDTG8MS`<
ztDN|peR>@D8qd7WGwS~b0QJqjUk3-<kQe_(IZNq%b@Vdu0>|F~egH_X-{zWfybgG7
z`}Rv&kcMC8+HIWw1*QB~pR8LNONTbc_D38y0P#~-J`8+?`|?x1DZk3Ly8aC1PbtOM
zHoprz1?=G&-?i-#&eebQ_9W#KfboQMkxusj?{Ke0DUZcP-=l1*{~x6k7yD`de}!Y|
zY2EV6XQk%|=XY@ayBy2wM*;gJKFW)H{1L}@P#O!|4ZOm+Jg5QkLAl<B?k`c7w2_8;
z0PFUie3lNzm5*`$AYgr`fKiU6&DSYyPa6LR$J^BR7Nzp2{QCYgJa>lj8s*u%{2BKa
zdG=|}wWs1Luas%~sGP|QX{}6)uR8TZKsu@a^1lbj?^6G%)AK*!8D(FYdWuq9r19TS
zK0x_#K)I41b>I;2GWTp#9F$S{f1OepR+pv$d3S*FdCE5c`|PpurvBeX`3U#_JD~j$
z?@v*_!o4j@|K8-9_vOEMDkt)J0uc9AO6hG~3ml7wb!!Win~wnB;-36hKW+i!$tQvH
zTzigE{>vlf<0q89`#Pn1^DM9e$fxba#DDo99?FvaQ~$e^z9)@-%K3}n@Cv1JAx*^X
zRgSw{t8lzdxrcL~mH)R>djB;_>E^Y=fV7a;lfKJ&jbr(vE!_s6FH(w+JeL2;igf%m
z$Lf{+@!q?Xo4_@|zA1x?y#H;E?cbkM+K$(^x&HqAnzqOL@;?JM0D1Nj*X_G9s|`|b
z<-dF!1s>$uKLhen{kKo@{|wJ4Gw)DloGZIurF@7|+~ofY+%t}Of%7*2?ZRD@cL37;
zIPeU|@?TlFljDB?u6mto@>hE$ZwvlQ8*$ykecSptuph85@=^Y%)9Q<K|5Hl&|0Bx(
zM)_xyw^1%a8*$Liy-i)pj&wdsX`AZfo1Fh9rG1nC4a(mEq>;R+0&nxoBKPF??VR5V
z6tb&KTi;)DPrGJ6d|q6?0LZ6n;31FKflGjLrF<)cw{!dr;0*9K&;5*2U3iM}8g0tY
zO(5ggJ{;lLzt_31{$F2AxGHPbC-1eF^MHDzyuHl5IZAEmi_{_R9|JxFjB@X@l;5V5
zKiibo0p&osRt{z<<%Kl5!ug5(xnH+LKIFf;B>z2rj8gmgVUCqS>1toJ|3Boq^m~_b
z5>STKb8VHp-sGCHqbxiQ{2Cyyq=$W!e#$@v_#@y&u0230|GjsecJAYv&%DI3{a404
z{{uk&tEbw$-vIuRbL$pA+gI0x>(KQ+&c#(4*ne>zp}d=G(nec%2=Lu6a_?J|^7~%_
z{en7Ye_rSMBK0X-%7t?EdBA>t4N!KvoPP$8X7)=v^(r9$Ehni%9r}BY?ejka^8YQ$
z_fsl&w(IlCfIO4-%J|D%-=h30N@Y*n-=Qw`KjU2aQtqVF^OSD@MW3}bKJTA$evNZ=
z)^dXK1?rJb+lxt;)E#yFI_DqdcpMnxT$(?~@pn1CjZz(+=lENIHdmSdCdY5{-0ySz
zDd2X%Ha`yR2kfJK{7L?fbW|T+;9UM|+thu1yw~LaNv^K~zGwfWt8~=9NOSw@@hg<#
zu8fG=t2|@B#Q7cWiHo$hw14vaHA?Tv_rqMbw7n6osVmCSo4{?rHR$nWN_qBEo>Nz!
z<y`)&|33!omwkGd@;vZ5&&j{XfP(+($(vkP&ZM3Fe4DzT<hs1Si}Qm2((wZDC9Wxp
z4*=GyjkK@6V}F0hHRZr(<p0}1!T%p{tgMVuD$}o0swdl&^1smkZE&!!PjF2f?DOLs
z?*)`~ZH=`29@qB(S2>oxZ&0`Nm;Ybk`X=y89N!CQkF+uJU;CvziMKeP2Bgzj%EJD8
zuD;5DWo!|g<cICbf8SOAl_{UQ#_>M_wlBZLMZ54#u3hI{@l!{x0KW!^oBWr4%G+z4
zODkzm@L#^{p&sj&f3H&(^7ubFR?h5)cG7m`zxpp-UIuLc3xI9Q5BVqm7r{rl{XWl3
zQWpB}d++DIG9eA`2c-8sfPQopkglJj{1N3Hz!=BULjRcoZH~oTUVWAF4ZyZvqSXJ{
z*Ab4T+h0>k`)x}3uO3W#4t|qd+sE-;l;?q`IhOvmW1o)$`qe$0t9xEQMOpCwBOKSb
zc8$8^zjR!uRIbIlM=2krwK&+%yMcoLKjs;qlQws7eUW->-*$aZ8J7RrL1jpK*w^3W
zSpHw9J=^~Z@D9h?pc8=nU*uf*lK&r|{1~OO`3%RifObgz{~5>9N<D7?>Vz_O4P2zh
zU4Z&8|E23$K>hzZAiei<F8{k6%kLV;wl6Kd&9#dh7yN&WW9#4Kp8S_i%Rs?@-_>4x
ziD%^hHqWR>^8XEvrSU!7TLlW9%UAoO{>yJ=NZplouK?m%=)e8nreDg6rMT#W|DN(T
zKzzl=z6kpN$0)UZUOxyt10L_ARCdKldaC~mz^?!mU<P=K>&nm`K%N);|18IU$31EJ
zV@m6Ki&FkeCuuGn<bTF>`LDcN$09f@E0*$K8p;>@`YOlTV)_5J&v0Jw-*=}0W#Ad^
z$)_V6zd~t!8DM?VMVkB*@DIS>0P<fsu}_}?e#kZX|F6J{z_WlpO1RE*>a+YG;n;Sp
ze@|X2WAmJ=|JE)4g@XUZ`ODlB7x6pK@dr45kz=o^2hw;lf99){;w=A1xGpcn_jXF<
z;Z;g?cAHW;vd(!*pS6xc|F@y3wnEu2Xe9q1<lMTR<#>})+zS1FgSve78t;CD`yU3B
zF=?gTJ_l&Gr2hkeb;$oh|D~Do@qJ2heTn+C+tO)}|GsCxeeTPE{d<&i+gDGd-v~G;
zpAP}=@QgGo_<twozstF@dyQlHBL6MLG2_}Jl-iQl0ePm)`7=Pdus!w4@@rg^|LWDZ
zfI|P3<+r(SKjgnWw|;rL2yTDD{l_?di}L-HZ*otXTn4ls_W=6;9w3dkDCL#>ALUp&
zO0T_?%CbE7_<2h4m;WyTcLA^R-YBKGi>p5I1<DD^YuqdBzjgc<uH8v_6p;VgK;^^d
z9`u-N_I(q$k7H^0642yao=dZB`ltT;oci~Bz*Ah;mW%hTfck$GP!<aQzsoa&`oBm$
z>c8#Z39NGe-vN0c|Ha`Aj&1)HO8Kw;TVChg`#2ut|64qxoZn6<@9yIG8IH#(4+B5t
z+P8fMu)p?2n#h0YCGDi4&neT|)xYGL@HIgEwK?)ddnvzn^8dSB&$w=#%C$HZ{Qo}3
zALQB^ApbuO90#`Pi~RSR^{Th_QTizh+C%Hp9=uBFvAF*Qa1;>Fdnl#R-vZKPo3hY<
zW$ZS<zFwewit;u<{vYDpckPS%um4wWq`mrOU*ws3^G%+Up4Y%vTy6o>{|;bZ<YmEs
zWnKI$!0WuD{r?{2t(4+_2c>-dD5Z7lw@0}y|Mj8Dd=>Cn-;pokVt=3D_&PXz>d37y
z^NwvOd^<hqcI(Y<qt(ogcd|q4>)GS2UUsTpuT>6pRvVqw?9poHa<$obU*&;TYh}Hj
z9d1_}%}VuBy1cj&S0|QR8>d>;jqKcN^=iG|-tlykTZ2zF)*9KFYI~#FfADO*Q?Is{
zS1V0!ZVo*BP<5@fls(e9y4Gsf8{M+z@`YaK>cHcb6YF;7p;oirx$)U`Zw@?ts8Opn
zvr`k<1Fenv)%R^x9%(i%)!Us$cPndM%+9ykd`H=;RDEo5xiisC7bfbp-v9H*y|rrR
zYW8ro(Yc|+>v3-_ZVfb7ISvBV>`ZID)#<IRHk$9dT4{J`Vza!oxUx}gtV__!p`}i@
zU0v>0&ac+9T5F@xtTww@r~bKKy}4Y^xYz8|ddtDR_^Py%RonG!soJU6GKfGOo9p!}
zS-Za6Yqw!@)@oPyGW1ewz2{5S^{iR%wWH`>Znf7sS$DPC%^J<+^<J%>K^hw3bJaG7
zwXB-e>Tx!aiO^EL)2(dQ+ZS7HFk8tA4Awg>T5dM$%U!Cg4^@#@!1Z>kx3X#r)j|L(
z8};Q?B(1YCkzJ@~o9)(Ss{<0?csc7d><Wj0C*53LZFFg&*RGF6E9WJKsMOnIl|+F~
zYrVb>9|zjV$~HP<S*_98tT#L2e!0<IEy)CF?d|}fIu&)KO>Gm|u`BesvB8CoRfA8q
zQ3ac3&XDzL@S|JpK+eu;>$2$4kDNS6(^5U_^sJ=ShSb^TdewF1?ec1a9xl}rE#c#4
z71q#&#?|T)cf_d;?;CBO=rmTEjsNxaKVC*8*0(C_^$yGft*qNpFdEH?>~yc4TFu0-
zeyQHcrspSSv%@EjJrbzHoC_z<XVa5Yv)Qrc$~x*kUg@^RAE=|6bWM^DbZ0Ds3GLP;
zn{S~diMv-Covv3D2k`6k*1OO)i4fPJT&KPP{#;yw8NN3iD1abs?1lp<+>f{}uil74
zrB!x{_Mi(m)2~8E=tO-yD4lV)TOV&-9H)u#YP*}d-C4$psM@nr6O&o>Vz=I|1i9=2
z;*-rx%w@IeR`9moSXu3651e@f+^bn<18!s&RY)xib%=cV@^Y)AM8s8C3^8(!Uu?HF
z;!(I<Z&wxBiON)VyxDEHa_z00Ej4t9LXKtYExEngYInzZeFF*JXtlSpC$p*Pr$Z)I
z=%<oV==hV_{-<~9bp<&?oVtzW9^aOaWvkUoidwyyAwum9w4_f>6dl2pdPN_lUUd`J
zBS#RMx=I@qeUx@cr-bYqBp!$B8&uVUkyN=}ZLajHE4e1*@9e5OPiE5x#-_*So}S1~
z^-$<m4+XpkcOY<Jj?SW_e&gO~_UJq^zuCiA$&jU1uUV_M8z2fZ`eYey9+5$va_csp
zo2AO2S?Od?&P<JE)4<fs(-W2V9aB-G(RO1g?B_&wA{YKvTO--%WKaM@J&`4^mD@16
zv$<Yfw$Eozo$4fW)2Xjt951b3-~-j$FtcJmQittY{Ze(kCp)@0<g8JvS0}RMoX1j4
z;Z?9_b&Ue#uxnv2wNjFY>f|#rdA@HZ3;C3DtNRHF;8BRga=W+OK-Rb5G_K+jHNbN0
z2Vz&nWMbDt%t1z{@ykRJFiP)V_P4WKUFWr0t2@p$rMz$F$FdD<L}PrFjSnir{TETK
z-bNi!AHUeHbK1fgt8b7H>q3W;eYKBE^<^H7?zMQeoncfvJUWPWlc+-R7W+!A-qp8e
zrwl*}a)nt#QiE1FlU_l>g6_J49#&Qpn{3M^<O{nSs-Qt?wHqrqXQiW7*ZCxM!1T4+
z)5KR4$VvnzRo!E~^dS%_d`}a@xW1!yxC=|@^kxe$r3=tN*<PawO$M67Ji)~JMg2j8
ze(=D>x)wBfjZ2M=KIUprt^+11a*>270MZY32YMMcC#+y1l7efWQN%DKp;|UsN|{fj
zj^_Ox8<2TXmJJP9R8VLvTVdx5T-Sh8$-PI&??$!LAr#|&b$PkIuHmGOF#n2ynmSS0
zx`6!2^>F-<GF&JT%wViKQMvjH9y?oBI5UNg9BGx*Lzqsb2<dp!Hj?a}X}6XU2K#(&
ztJBrUjh;Gw=A52%|K73egeLoN9cww39XT|XJ#y~QMBl>aZh=iZ8fXa}yog7Eu|{jX
zLb`(kszYaMW2v=XszKpl;4lvdrCPl%hn2!4!{0;e1_yFPeaaPx@SufRuWyFb4mM`k
z5JA*dwo-3OX~d>);?echMy*zFR>CQCR=bFsF85-uIphk_&r2a9>G9l$taeAHlr(Wk
zNGrt!R$13zQnS`XfssODnbeaI@MgG6>`FZ%(vUuA7BNv+1(gTmRC?k1m8yXv4jWm(
z?_eT`>^s?V8;yb7n4$2fxc!1)S=u8-ACG<~?}}?N^lqr*>y3>@m!P_9kcXpNZnrQ5
zMMD!AoK<@?g?J|jlY*^UWiUj`r97<6P6mKgsk+Ns5GuvSNi1@$rH74W0!rNed~*As
zSPRQoT_>jRiw@3n-_y~3V@9n_KWa<XE7zB*<^kF*Yp+t=f{!K|$5x$(S+KW;dBxz`
zKtT_m9AMBu%S08e%i390pD|3cGf_EI1os1hcM*KZ7z7v;y7d(YAqFT`Z=H2UMG(8>
z(O{gBqC%Vs!K$7iS17tt*r*7;;ph5#>$269Z&Q1%j<G|UF_g&>qdy?LQf)^ZftWVh
z4h_0FeZm<h!42!HvXywrkm<3heMBn98D<b45*er(3|XXW+@FRWC23MLv<q1XKI&P|
zk()eeH9uQuSNMl2C>8-;J_wqq&}4)JA$pfl^?b;rstq`L2$6f1!T5%Vk7zE)Q&gS1
z3pm%+H#^>udWb-bePSfF>tIULQfpTakXX2!t*p0};N}pWml(GhZSJ&M6EU<Mz%}u#
z6P|8Re=w7Ty(u~v)PF4tLm;Sw+R@J#EAEV4$ycDv<yfW|a7FZwg|LsI?9g7@gbov`
z<L^BJZz3T0*K1>Y34>xxaAs<9G6n_{6$Ufa<+T-b7JsPTthDGog4QNU5Fy|0>XLL*
zr{btv->$w54lRyd!K2ndDtAE)L@G~CPa;i+x4QLH^{&=nE?I*kjOx!#pX16r_Z37i
zNG1@OQl5U!pr39Mv^16x-p0CnR8^MNTgz)>xl58hxk(=m!orXhqtrw?8s5p{($?dd
zLSp-);jvJzW{<tG+P4u#igXL}lmZnf=x7%LAYKTgeo2|Ys4)z4P{r80+FokFJVMi*
zI$F4x4v|Lda>pT%bS_o9kje5n##)TxBV1Gr3X&>f>n-d9u4bcf+1$^Wko&KV>RO#P
z9Q7flwMN8Z#JS?WPNmJ=6(V8N?bAq}P9<})LyeQJH#+Dgf|aOJm9v_vZkDJxOhF|9
z`8N2DzB@)OMUxsBV-U>~W<g%EB{8VUYP~v1fQ%7hyGm4x$1AMW_~np(UL2qq%9pCi
zus{(g^B(bjn5~l$=|6}y$@;FK+943{;E~x8O=H&)2UKfjVtS%-ZfI0AmUZwF1}M02
z9ltNY?tyU=u>&ZSIQAV)5*a%HHzFJ!AaeytBgWw_YF>I7kVzvs`ky?Tk<hTvOFubN
zyfl#wqZ~1S<ETm)bL_pb;>BJXPf?vts-E(TDM?i3N@%I7`bFFA`l`vPRud-Xc|aM}
z>1~iU&7%VJhfUrbmg=S(>w4>m2}9zn$ri>NSdrxP^jx@GJ^yg8a8zt6H$4r^FyKAF
zfcIFlHclo66I4V+x#NN0h<KT09CS^)u_1RS(ygG`c!l9e3{`iyo_=1!%oz&Hz&=Iy
zNS8(pX^AlhTjb^01d^d7ptym7zf4k1fyEM+_e<ZOi2daBJbgP)k_;Y3nns($0A;P&
zx{Q3S5DJ&hdZz{=YhKF7(&1Aa+m=CyBB#qN(}jWXkVM<1LbP_W08AP&s2mRMLzRqD
zdNJg<k%*B~z#<J;`NGs<m5U%KS2IK7I#oi=LJOl)=lhufM(8c90``_5`GzC~CTqDz
z1T3o6+9hlcNgaG6_9BHD{gls>QwUYzqMnrSGucVEKg<pdBycMyHA@lob%|&Z@;nJv
z^t254`jpNSK^mWwp)PVbj5<hPm}J9}&XGl$R+b-RY?1QB`QzlB7zput@)~7J3TYZ~
z?PSjAm8p9d78dTkk~R0j!~OtCQv&5pML9m+`QFuo+33~u>Q!E4s(~Ihv#SZ6enOmU
z#po*#+cB)mMz>NW1VVdDEDcFW6{YNjR=yfZV1|HQM%+Y14)}t-p$2BCsy&pNco&QB
z0K4+!0RjfSNH4PZ%6Bo4Bj_qns^w(ieg3hLrL4g2B=TLxR27eg)uECmmg&auE{UUR
z1S*P*p~_HL`qV8<?<PU-5?EpJcag!(7H(ktGBYn@j8ibJj!>L&;PLh97HOaeN0vwn
zMehr)>&tVJ<66SR<YB>j{bDkmq;g35u2zZq!0=*re`7^qni|9JM)C`yF%M9gnV2GJ
z-k%*gn&w8bb7I}Wdd97V7)lk*(`Ut;X$|IQ!U6fPho=r^`%LX2MvhWbADC>Ahj9{a
zg@C8h=`CUW>|z?G<;F<P4kXrg60Mx~7#lIpk`8hjo4_zK#3DDWW2J8&uS^$xPrVqh
z5ZQQA^&!?xRAflvhZ2?Z<06xZ;{;Tf9bpw?7iW0tjMQ7?cTky5qn0OEeMx&t7&J*}
z8)0=4<<`ZB77hCPG=TkK;2<co3yD%q6E#J;%Zdo6&d0nKw$ENC)spN<M+%BUhsp^_
zdx{!Hvn!v-&h?g}PEPic*g?(_g~5xE-5SeONSP2z2nCIc>WK~&Wsqr^^Jy^1NLgJ2
z;$3K9Mgr~Vch+FZ8bg=4$@o>L=qQD&%(A2%X7PQB2b_%5Yti43=9)8CMip5tS%ib`
z44X|<a`8=^F07J5m8SD<Mre=Yi*pO1w-dKv!L=Vj`9d!O;bDVF)KE9yk3m~)AUGuz
zip)+P7m%{qEt?$P$d>!O4jSbOOUS&uM&w3nhxw2K%N(6>khyT~#ui5Ej@1gOj1HFQ
zLH1uDkukd&WQ7Bbiza+j#6dsXSMq}81xEt$2#ZLz8SRykcp4H#02VPMln1Hg*ajX;
z!<7o@)io@24Dps2_2g-2b57us4;OOUm4XE~5dYT2iyg+EPUhiziuT8F#2pN!vY1}4
z;6`eV(B@=~i$^m|TrnJ?4!jKFC=m?7h`=MCNm~P0GlU}Ipk^GlrgyqOoTwZcjQ&eS
zv<W6!ix4L!web0y5g#(=8j~3?Q9mAIqP6<gxDzer7#jEweSCq;SS4w`NXDyU8IQy2
zctJg-rkq-mNVUC$A!`*zI=`3O|DJA2KZ^e9Z8*dG8b{bbQD?u&^b~L8^E;px>XXnN
ztQ4j=B;43!1ihqzA-i1Uky7=Q)b$vh*&6eqFpU|85E^<F<i)q0!w-SjQKoRncN^vo
z&>ZXSaEv~a$}Q&=2S#Cc|Gd}l&PH?3GEF>m9xPHAVH85*O+yuEnxT`j(pqahV7y39
zcqh?n$GuCTVMUN_Cz*%JjbT(SnJY6t55r2f+sL(2V7-uy8ZYeqxkQ@CMkAKkyK7xK
z$fC1|U_?R#vth}EXYF1y68RX+BF{20$U%9KB(c`GMEH>8Vrgr$mJ<kxN52sA*2V5d
z^-3-cdx){rxmMO%V%(UtFG(4Mj?+paZ6p~tW6pb-SSyD3=9<&2TYtFMmol`mP$@-C
z9x`OZMpv74;^Xx#{8EbRx~-KuoM@AoQP~K|NpfGwKAl7&m_EScq*tV~cf5`4@OHe*
zK%3<_&Dx`b0uVt%UNtgQT9?j3KRZ2|efnOfF(1D7(QI#au%OoHryo9Ffs}h6%|@s8
zMsm;SV<R7~Y=rHKQT~mDNVYQhK}IzF2S^y3p+k5FXGceq7prWpV;mg=?d`uZ5lWbx
zxv3n+?-3%Q`2C1cUZRZE^%W)~NmOBvuPoO$yAcPg15U|m{L-`n#%7A1F;ECkJ%nMd
zIVq}m4`Y%dbQ9e-26r)Dtsn(7M|LGo9fm~1n3$pyAxSAyk&bY}Nes!Jr4ZPLS-gY^
zV`h7|d1Rxt4c8W2tTM_;OOzt{)<Ht*<K>i!v!Q6e%Qh5YuyWpm0cY|MgArTm)haew
zr?g4li7oeQILi>sU{TH>l*IUcqLTV!zA&v18L&ueMK8<XoB69en~`@~ZP-Od6w>q{
zsj|N98sNS5o2U9*1%de0fcBtGTzit03x?XAUYfTL-z>86Z2!@U_Cj<AT1$D;{7Fc-
zQDsuG%epWWn8g;#QpnYiWq~fCj|?ygG6_&4Vusqxj8zyBGBf0yZlV>lJ~gMSJF^iJ
z4<rv`c$m_PVRRJUWO9ZOT-pjDHXtA**wthoPm>wCh@7M;0?384uT7-|3)h%O$K$2T
zVNjL~ZO&j7IJa*^2YLElv07pZHRj=Y1Enle%INQB2?>H`#f!<a4H&uAl<G_<6j@91
z(k74*#8jEC%&LOGGy(5wYgAIZRANg=qF48!7_GY0Dnf=71Tfyn>%ha~dK?f8dfgP>
z40>342(!_nTz=*JF0@nA@>OeN*;b<-vk6)4(jXm~!;d+%+`;Jo!ADDgOT_rqne;V2
zg20?}XGITH;u}3P<VXIvE+UZITtuo6g)+a)<hr%+MB%h2DoH?GcQSxGE|l4Us3(nx
zpy+2!FX%V5wK4s`f{JqbKjt3#IpyJmkD+@Qpa@|5PZhy0qxK8vu1+$U$37KB4D1aB
zBJa+WI6)i{i`p<znLW=`?)rMqiD~@r<<&HD7*YrvYEoFKKujsu@?c(jO28h=7B6O>
zle}vZeK0wkRxjcDS>EE3y_Ce?sADN7la#vd1)fZWj}GcPTiQ2KW7+4jTh<PqsjXES
z2d6&AIvW?8Ij??09eAJ0Y_P=n0bOw6evgD4j0U4p(kqMO)x8%^=aG1vHxBkgQr^Gt
z=}P0?hc7(L|BqgHw9?p%ady0}hzzQXy;e$Nl~F%(@b&4lDE_0Yg(63JD|%&RajiI~
z+Wz^a3PibZdiv~Y?b5Be#kg@+HXli@9?a%%HTO~|$bbwM19l9v9YaR$=WtkI(Lh2C
z4E4d<rGu%lgVeciHX}Y@)PmUB(PDC_GzD6|sL7g>k?*#c(?{FNMFVlhYF;9W!W&0*
zhz@TjL5=lV12~6`GWXUG%Ly>L`S@WbJ4ECeu}m^8eNL&9hAJmb1m7_KOCJY^^aBls
zvb!X#uT+;=sJT%2;PECuld~fYT&mUGC`3w!NS-1jrSsVAVr6EL@GNqJ>6J0r8$``l
zh3pJThbzg=TC%W3T3S7Us=1L#tI!HlP?)b`Y!AtEgA>^cWBcOKm}wRN;dFIg`v!5!
za+t76zN)BbzRZf3<mk_jMV&?B+nJu?ct;k$lGZ_vF|#~T`QUkwDAG3AYex?V%~19t
zCwN*CvFp7UAm@Y{^r}Jd-C5LerJsu^2|q#-9QLxpB1!Zx?MrqB=W{YMCleXjJgHhx
zwqQcpC+t+t1^KZX4|Y_a`cYP2`QW(}&>GwAN|>U_g6|<5Nk?~LRXoxc4ZlX%4XNVE
z9_fDYMsWMnis1u_sv{n361MxCT``N~Xx98JL$SV}`(Iywp4IC?Y!fLd)tz1<H^-1c
zf!r~jprIv3zzi66ma#RD60IXDj!`u!DY<~-xN&``nWa2E<!CJ%B|#J3o~B2p7@yhy
zJPON^3#};fC&f@Vt$yg|H|#I|la)T&zQ|4+_BelA9|#g=CT93RBP}|O8HVhB$2b_|
zNMN(<!sRv{7vBtKIr|DWm<k*);ki~wPpw}eZ*r=jZK+SLts9=y*lnCuWJ0ws8r#Iu
zXtH270;s#`8gqwC6f+Ed$)IJ5m0fwxB}Qa<h#7O6r6}ZqKs7sCKk*0)++uPN;@O#e
z8O229{`*gjUUBqHo?&z{yZ6e4Z0ZwBQ+qE=e`0Am8=cl2`-Ov2^B%3}nY|s#o-efP
zTl=C@1h+i2r8^w(1*J#tRKl(&-G}FKz(BW|)(GP>l)yYOm_{*BP$FVN8I!_=Vy#Tx
z@3lJ7H;H%#gG4^zvP$<FKF<o1WtPafF`~viCP+DULbBve%)anAD=pJRP6SB96;^5&
zFCJt$>FhkI`5C_1I94S0L6cg}r9oGvG&Kr$>R)NagM~xW(-1C25ygO_KW9MKQ_s=e
zU_nx;ZflLSZZw(qD!;<hd4e;DxpP(_v_Ep(GD(*9T!LJ-yMzDZ6<7Z-%vPrCU!lY0
zplk<rg~pUfX~|4BqXSMn;?6)^W3*I_JSb@>{Zaq(ku<Da9g?^_&zcYSi}gWbuUwRA
z@IsSwcG~oKj&LzT8&)5a4&_7x`JdVT`ktMY23mQd!T8=Upo+wtX}vT=Q6}a}D+OJT
z>D!QH1j6uK1CyRCO-j=xRA~jFV=8FOl>t!?ubLLWp!UJCayaDLrqH(B)uqSM2^I#2
zO`|#H%;YzkG0=784#%YtVHu0&Tyn-@E=H!TwsCQbE01vQ#Y+FYm0Z$t6JIog*)q^Y
zxfq#6n^@uO?&4)=5tbENbnaHBZ7!*1NowRd(5TFvj}NP+7P6DoE95BofT7CJtXvO_
zBg-|^SUDx5f^^K3$K(AqR|&mXv)o_6R*{l@`;}WHeykB$<U6;(t>{<YD!H!w#_+I`
z%lX6|WAprpWR_TFj9G$a1I$bs+qUn;U5}g2HY~N;tF2ayj7br=tT^T$h9bC$G!2@a
z2Upg?m3@aG*$Ja*e5PEF{uZ=5BkfMa?@d-t^JH4Ymlkv#DwCF{N1M~P)b?hl$*9JE
z1K&Pz@0QxMmz(Lo%8|Y;hO?(f*KR4Eh__jmm#y*e-kq(C_Pb_HZC^{xS^2?swztCC
zfNqPmm$?}4=8t2t4PKpxcDo4}hC7df!eNyx%I$t-=fo`bm;7}(A`fSil8~<Is@I|=
zT%bmY5Ae3K<<%BB%fNBB_Q&|_FA_r`lw?8pJ<%+d*$|wBju@(hl4yn*$0f~Y8K8%S
zagF->V#;SXq0yMkh02*@*(goln@u<N9^|b!JILHTf76YFEDduMw|DTlPr#akH1S@K
zMoN%6b-7Aq(X;48BfnVEQktNy4QLmUO?jhWT=DvV<`dKulUm`|BG=OEv>J)Rd{#qJ
zECe+sCO1gNtkhj+h%)Fi(%KkYQ3MjxAMUoIzqd?(Sie?*E8)~YwICr;&~jB8<6l>|
zHuc0R9$Myy0ljq+PNtcnzol8Nv*wn#EUga3ROQ2tr7c%p$0R^W8zm<!yB0;s$eEqb
z#jcIPCrSE=wLGg`S`ToDna14XmIyP&mfj(Sja2OZRZ_}_7eY2j>3%rkieoHnEE!KB
zPOM}e67x!#V}!bC@oK)%wm*S^#*)>WB=>^kbO_0lah}Ooyqk?4+sjNLh!(UDX%3cF
z*JDLX-U44<CPqN`4I;hW$W9iz%J<eL(V+Sc1I*fQzTUyYp{6Oj=$GW-RQLvrbW*KS
z>%}TVYQYUOSIA;9tFe`qftH4@T%2F5ZbdTmBI<VWnL_%QF4S-Lfnx5z3dfxmDCKrO
zef)sbhg*q-MFSI3%1JTZ?KFLn(i=~5-|sc%Z}$anm7NRT^7y%=7jI%HL?BdYXq6i_
zha~5f#=Tcfmm$3QH=xzxQ=;Sxu_vtWNFPG=Q5p(d3%jf*O+(NLGDm3uLI&7KN4>Kf
zLoN(WOk*2A%;b|A^F}T+(l1pfDyN}IZfbVVh7I|=z_vex$X~2(_*r0by=5vb$vs-=
zXWm7sPg7t<)-<S#Bx13Ww(ma~NDwfkmGEisfoW$&Ope7s$Wn~OLWIXj=JK&^vGO-Q
zcgbU^auL=E-G@UUWj?)N>YT_5vO9B28&xMM&^t5<J{wFW-&D%M5W8ZLtzUMtnuFKU
zWFy~M$2#*gV^m2IIniygGCrhZXS<Rom#o}H+tV=EF%UO|U1B0RPKwV0l0$C!xrYWJ
zu(hJH$LBeTkEuf{oOnpJ7Quda-PQ(6V$xuQrVFzd`7$d@LnGVIhaUwgu+rYwu2Vf0
zNEQ{H?2~wyC;58+2cdYlk8;5kmLi|Kz9BX(87;bo5jPAzhmg*DKfL%5Oe=k-pY#z`
zBsOK|4+aPM>`5LXBfUEnEWLs?a>*WYGL)5y?C`E%CnRDSIkIZAeuXiKE%k$DOov(f
za9VxCmm15}pc4^dzoLP+)4FHAlh>+Z07|{YPW6L$4AQW%i0bJB#yOu&4)edG{g}>7
zPN85q5tu))ZZscrnuCs`^ubA#%E%?93>>_~CyKlpD+(CWGh|hO;)^^f$nxv=SgnuE
z3nM~^I=st4L>D2dyX;C~7?E3IBE#DFW`l1R&F`j<h?e-QXx!CLESj9=!=m917zh_1
z6DD7lf2|5>3HM{1rwr?QiVqw7eB5Cv%1EFSI~k<l(l0c3@PO%G;v}qQKXzAp5U?=Y
znyFZD;YVHr;}|V}<LX=FRCtpwz4nzVN&{(Vn)!jIt{Vo&$~_;Um3)4*P~$>p3^`#+
zira{~6vmXqAhYGTuQF(JK`|dvM1Vqy8YZfdtRL<++?)ZPQXMI}*g%1p(SccP{c*7u
z3)PL%(K<ss7fj?1-egGO-qS=}o`<^Js4>!7yN@+=Eo?NZ>MLqtQo>li23d9*g-X#u
zIC+#APCl<RDG;7TJD=7T@}m7&%l7VqM`;6#k4AZhPR7Fbgk5K>(rTX`Ek|a#%PF3(
zvD8C-AE?`JB>6_b1VoTp#AMKr@sF>=A~GYv5GEV{n!YJ4Y+%_tDvAn9qkm**6gi;l
zX?43_=ETAgKg&nhVoZkv->LnvfeEQu@}Gz}bXqfatdIBG{2e{=8{G0c5+awr5cSPa
z8K6)BlLaA3F~a>sjB=T&@JoR(rE3S91NBLsel?<^b<Q!d)*^wa2ZF(f5tZj}4f?kC
zFzWiYDvpm?@RR}{N<Rky)2Dl-P!rvD%@RTSGWisaBr^W_ng~7|-&N;hp+I5*E!Nv4
z)ZE%7U+9dq68JO1h@B_;d3HaYD|X_@bG~KMh%;TF9t~(57hkOu-)$E{LJtUc8doOR
z!Gc)M)u=VKH`dchxU#`YOwDG?4u=MZZMp3mNc4pD<0Bd+)B4j2<NB4D&d0(LrMUM`
z`lK=DVPs8bi`g_6Da9-po96Ik4dc=N_bjkEp38j#j(X0|d}9zoQW{nlV~t|fYZ_n^
z48=1vxIz{w40NAT)WT}dqMqz1;c*b`aQ?x&!^5aSFP%F-*7C8%O6|togYun&_>L8m
z8D_~fj$yiy(a8VJlDQ-q>Sh>y_ZCZT)u!a`FuqRQr`W$nfNu?b?-bX)^kg=Jnk6?H
zpWYMO;ls@z_u8xhtFWJnOCm0{ma9wdd=shvgvZQ&QW&U(3GOuCehlmD7D#FDA?nS)
z#}1;qqk_wNV2aGQ-nRX(4DFtjZ(Fo`|7Z4{ghNFS70?c|Hl!j`k{!_}2awH7i`RgA
z9zdS{`Y}U6CJdn~m4jBXKLI^27F9Zn7TSr&(g&5kFqzeS7PJrzcFbaX18tZSx>TcK
zNR$6L!g*Wy6B2)*b|^AEC)$bbMUU)aTrc>LQyYn!&K^X7h}~00>Kvb<*Lv&qAz>?$
z*o1Oe6-E?AL=zTqK0Ed3@uSBNW&LllW7UL`9b2&Ekx&Rv1kxmCur0;H4`ye|k5Kb1
z7mD=sW|ar?d{aT8{7b4<Yb}@KNoHlz=QClT(k@am0977rKQcB@#3dJKFqx#Wpj7O{
zffeBEi}<)a$zWjLC^l0Y8|~wmbT3a^#sr$xB|AC8w<+mMFZJ79W8B@U#O}N(m6axW
zDO6hYW(IMpj22s9?bS5QWT*3WkZqY0ai?88#NH6)5=jHbDR!NbF&~&Zm`yGm82i9<
zl=BF|2Td4`90)reU%w_QMP3O>s>FJx@{0}PKu2zNE;n2ZK9))8Il(bQeh6X)m5n%2
z>V?XBRB*61GPy8U3AW$Dz<M+&@)>gUBa?6Z5_@mX-CJI{1<BiBs{Auv>5NP*%<aua
zXI$co72T0Nydf;f$IQ`aN~<CN5%oeBY=rwP4b@+EB$z~mq+;YTftsWZh%t}bk(q$p
zWUH%){-`+btns8|>EO*#ns!TXl{*B#@k>(?L1KAS8Ajk>HmlId(An&4eMMWCe?u%+
z^Ce0tO%j7Z>&9ISJjZw8Lus!C68AWFUF1<QW@?9;Fbd#Y%P%mDKq~Sc<b!1@jzn%b
zO-@)<hkCg97%(%P6J|`341EVvd=y)J-etc0cz+T;?DcT-dBY^vBgdxL4o(8W1luU$
zMmkbBmI!ly-p7M5lGH=)H>|g>@rB>OgeJKr(JDQd7@u5V_IQ69QBU9aO<lfmjJt3#
zUBU(f?$Z|=*6GI?RuIpm4R+pZ)R)_c!jZA2CwIsS+`!e`O|e-9^55x(3-f@miD-$m
zw{eI@HwjpwhiFEJZ#qw!@(<+B7m>vgm&A|Wf8z9+&fdzgQ{zXEW&D5tS^qDNE9I^g
zei1x9GqHb;ug>b5<K#@UPdJ<TNOO%%*Vxu><^T?2YGLx0smbZd$^)lPp67wHC)|;Z
zH)kjI`NV<gY<gjaN2U%;EzG7zj}ZFCHW=ijr-^I!A@>ur`?3QI`+RU7MNbc&Y#qy<
zWTPB5%bA#*@u|6eRA-A;H=Ulk5xryiZiWd@E6?U{#B<;&5#Bd3GwFtdpfj~FA5~8+
zOj%CF43`-9UlDcQ0QISf=>ysRh5ds_O;sRzvD0NX+Lyb%=o;2~W}awnW??4trTHBX
zH&%GcEg)m}!-9HD_ARlz;oJr`C6;v`9oF=a=wu?$%=LJ<Q$CE0#A2N6gX^6T@h4{{
z<AX~Um#i|>jWzG(N`73E>Jmme&3r_Pj8T7#z<Tiw85RL+!r~z|n`D5C@n@4#c1BW)
z`$`e}&Q<nd&SKBW%04=NQXL{v2pTccG{`)ig4P>$y6kHQ@?CBYW9**NczVpZqn|+?
z6V9W<<cFeeERyT=-bf|nEBB+LvBN;2w^&9}D%f_&7_~9efp=lARlNXlDAtKpw|ip+
zlom0x9;vL@YXTDVK6bqH@qxW=-g}qJ>G=iM)<bauH(V2+GM9PH$V#M6!73u(JmVMU
zEG9?GC?j=mhMEe6?-~frEX);KWzG*oVKHhQ4#P0yL|z@&c1E5sH)IfGlP_%D!?UD_
zOfE@*=l_KUNh{?@EB`*&aWS8exbQtEvf2J%#%>~zo5CAKU?>tRHuVDi0SUC_S<DnE
zkMe_!U=69z+;k4GzTeY4&yz$Ty4e4`aq8f#3;>@rX4B;6mH9#C&$IuD(KpCvF1S%6
zWHwN_(G^6@P`K^zTLVtZ)C|(WvUz$nxiG_8FDeN+VVIXwx$t{2c7O`S@IC*;jFugd
z5X7X#?Y_OvhdbFEa|`w5wMaIQ6pX<fA5<Xym-&#%lqe!FH91a1=PHp%5MXcyeQz1s
z1`eqW+AEuwo*Zn$7%A<|vQ#IAO?wHZEK3#F>!o}}#jLWJr=iFaRM+{O@>jKS`*cJ>
zZo0o?Sdnm#c}h@^h_1igW3GHj9BZt-TyJg0rY(%NB6f4sCEj_au}DqkNTd+VMVipZ
zTq_Yy%JSt8d?K5hOxBYC0*X2EOX&)d>I4w%To)Tj2L6@YYWeYL>?wnGD)nSdn2ANy
z>Al)SyqRtcDoMGC77q;xYuNL&6fFdN$EK>~nrlYB&c!hs79BTP8Us~(-(Oeb8dQje
zy$b;u$n@t!5rzV)N>)vQf0$0c4j~NnEo&>$^W&Oi+xrc;yKhv59?}(}tP+!4h52?W
z4T@Q5qgb<IwZ7blZ|F<x#uIkSRX#oQqg!5$AeY+bCh$yroMh0%X6k%Ep;zj-1x>NR
zXkdxE!!yL!6};V_nUMnwe)I1(?9cgr01yJ752<!oTxim-#D(%Cq+<55Ad8!Ar}@0W
z>^)K6fOW{jM<`f2NFD$QrU{3HcZ{Tb8BF3RilO45isAeF!&%U?FMX4iK4?n9+ibDO
zqA&`5qY>5eMkMuk8|4dGU1m^-yq10#19iNa`{Y<ZfS*0&_6;!)?y|Z;5i1jX>D!AC
z1E1($cP10z_|w}Md!|^YoXcS_w?+(G&~*1HVSFzl)J*$k4w2wo>UC^=HBI+ofov{j
zrW*&o2xee9IdLF)AF@OAy<DLXK?asXJ>df!hGfcMOeAX39sIBvoL?Q`wo_U%eP&Xb
zEBKVJr$+LoC+5e%Xa876WE{jeK*sL8vHiJl9uLoBVjiC5zmv1+0e%0D56&v{dCagd
zAoE53a6o{Wks)HscXFS|9%<%rNRE#~-To4h7|*%(Cf7`Z-=y};&Es>^3LXp>ZzVDP
z1+x4Lg!n|nMDB5);#X(Ht(yzuCg@`NY#=$IISwr>=h2D*Wil>IJUhkZgbe|2Zdx&)
zEt`lMt-5S$qH-$B!=jA!z^k0B)h>l!Ce{wCmYq3sB8f!gwK{Ck>ISc*jV_VUi2rY$
zn%ox~;<LewTYbc$;(5Ot;@>1Ovj77h&SqmsH8;V{xe5w170>$)`MLv+yAH6_Womjp
zU~Ynh0%NSX3d5S2Y5(!Sl$Yi!oH*E_#fizdG+P-xyf@DI&yiU+7f+S<vwf45(Ib1O
zk-90~KmfCueKg0TG;JPgZn82@6Z7-YYBoPvnP$wETApJtDT4erUtwI;N0Y%K4Jyi0
z`|?{<2^q|U%*|A0Cl4^viq|#M2f*U!-bo&(%Kb6ctW0v9u2HwW^&XXS52WHAHO1Nf
z3d@wr<pQP?id9mB@e|q|>y|p!8F=ikIMje#;vM)vqe;hdi7PWRc3_sR$Y%!ZTE%b4
z+}JRKvSHg=?Dmm9%OGMR3!ZkLywQ|JcvD+TgR?XM$Zdup4P${$Z-pJbu&7v=E;c*G
z656<s*>Y<MOp>yUS4m^Dc-*~4(UDGTldAlbJ5BcuY)Mfm7omj(8~Az<+q9)>417Zy
zHourMEqaKhPdz`v*C2RpSQfaVX70@EL=g*aB&%8a6laI#rp!vx#v<CK7co7(T3_Gf
zvkTow^%83Z?XOA6amGzyS|i<*uSV0CFqZRsBvszfAYY9(Bmnuc>|}g@@jAMMSkcR!
zOXS>uDgv{xgM&FIvPy1O$w<<$q+ELDymClJ?lbe4c)<LoXeiHFmQw`b?uWh{i2Ro0
zvT`TbzGu?%7rWalY&)k-p+8|hk~6S84H}1Ogu2C2{y|5;oQA6(VjX;ZcL?UgGl%&Q
zWFRidC*8zgpo7I%{$VYxWq2NEaIwcQ0{WI(1uxR?4r3lXc`>z$ey5q__#KN#7Vb8_
z97w<wx7iykO*D~#c#q+y10f=iw3GDUP`_|}?75Pby3pM`KOq5<`$2c}Xt+NvMK7F8
z-ZA<v(-Gx}J1qxAf!RI+zf442uG-I6E8u*~&}cl!RB9NiSge0YnOL~NHm${{DZwHY
znr$x9DhWQ-h!3oHFpW3J{g4YGaho>tR~dJ%XRujA9o=#oaHNQcO0^1ZVmi1AgDq{V
z>k+0Di{8s_41N^JTm7vgNKg<Z*BHyD4HE)|G%`x*Y>g!oXP6yEFBnZUm~2k^0M89s
zTC7&Au8DoVlb&hS80{x*ixu=(pelycMIi>a3S#UcY~UtGndv<bh5P_du}E_;<-iA`
zmjel-kKEi@=_U3Ajfr&3_-ePi*;)AHCs)v{-qHm6_(^sT^`{9gvNHLT=`i2CpNWD)
z%jJfc4p2Lh+(#_fv!PqpOhTtq=)l9(_2tziMx%Vr6t?&r5s1m^2U<ORli3D3#*(6o
z?OvnHi!REF8r^$}W!jNxP9$}XzmXDD6O{vJ>-i|RQ+bjs>O+l<jV-<}Xt3Y%7{9Nv
z(Wqx<w`l0$1{<;`%lSZ^NViS2ek^;?=~R9M=R|z2P|O*`*lPk;@L1!LV~<UyG`fsX
zIbE3AH@Saea^LLSWX3M2lT-U9*lqdg?C4dte`n(OaCMa*fvWb1eYbkel~w-Bj;yv)
z!;iN*t36Uf^(Nb1u#_PEW<W7t8TNENjr~c2<hu)`#rui|q(qKa&`x++Jj}58NbBrb
zaN!G?PfigROia%1p9$1vL2U-qpgM?=Cb{~5Fa#Ro8@FZ^!n5lF<J^<%A25F2K^{N&
zz|yOP>8E?G<?K<u89zY-C)1BC=rY$^iSUYC+J2L5Pf{|qz(~;@GLpj>?w_0?m40Ay
zU!ZvZFY~K^pgAk&s!h@-E0Bx-$FdU*mLC=PoOEmP^(uT~RS!RGav~Oq?fki$@p1{H
z0+l0%KaA~ick(l_SAH@Jx^F+>Q^`*z9(Hi=u$#`aiD`DS-e$9+=2FY4-y_u(6h3>P
zx4v9w?eG}ifRZlF4)dc`W<$oZ2V3ig7R*~;t~b}xjNT3*C>PgmuC`5+DQz)yO$h+h
zSl+jrH@ZUm=oC+ipVoI{#4zb+CyAOT_cL&FDhH}Fv2j1Xak+&RKuB4-i@SKt{eWQN
z0)oSgVKM!hf)keCc2{Hn#&RLiptmF47FuEsGL)ty@)QB_^vvA;K9QO7d#KmkVx^rd
z=l`+nJO3jm=yzUQ|IYuNNV<a&%Ki=t9^1ei5J3`2J01+W`IWDW>IQ%DDBmu&lo8~#
zjAg<aLg*a?o0^`OnVXv{3C5%gKN*J^e*oloT=gMAQfH^y-a30&9A(K)Eqef(uCTV3
z^SVnuANSkZ-1a<Pg_Hb}+|g=}$QR2)KxV8^uG243Lfy;bf(h>>!}}*0A<axpPhXl&
z1XiaGJk4;I-*p2W=1163HvMXc;#5?eBB+7jWe1IRwpZA|naTNy$@#(rl1O>dZ8-Uj
zJCcbj*@I}pV`%frRxf)(%QEH<{<S{my|Z;zsL4dtojo$4K!_PKpJs`Ru7*!&Y_JD3
z*(dCO%WZ`a%ZO*r_|qzU$;sz?MW}Q9*s)_0v&ebE-6OHC7yWxQ<>74hjNZK6^mDSJ
zMkuzaQ_lA9pZvl0|G7U`5DhzftY@s<+eDF<QE+pSW5eP%MveR^pL)a`Y!exX&VA=K
zR{qexW%h!v;Yj_p4EPZH7%};uzL+0aI*3qq7IcAT;|R3*0j`uEEXffI!yRHV(sC^t
zfiV${l8OBSQ+&`4Oy3)m-K}TokZ`@sq9;c*Ryi;_zjtA3|I8e;nc8nwXX?NK<%{3D
z%Fb3VqIVCqt}-`&xv|W4#i#lGA2OJY<`$VH^@HVpkb&%|v>W{O$3&{fvE-}0Ms}gc
zVxw>{a5Wa<#4oA@e+dJJe-mt{Tw-Sa7*CS>=V?sM?9_mNp2EEh{}5x)=J@(n6%p)q
z)(CoGQ0|W-*Bf-l-{$+Ip+4jakGt$JE%~R{ePMF|aDsp=K9V@Q!+?(-KXT&i-oD*)
zT6<?mgtv3rD?7&OMeHysz=v@l$WMdO5YY=hn>ES)a)~=)HnY>DmAKZ|?nKLnTjQ*f
z+2Z9(y|v{YKg0=9^Iq-drWs!qpIDRg-i4@%pxACQKL-wITwKT&%*S^;-U6kE8*CzX
zgtd+uJ7OuiVi<wOgyO3W;z@}jo}9?evMJaE+aKrKZ!>ZzKl&^zvw7L~BDER2_at@L
z>A()>a+IA8*jR8444eE|W^Qzh4iEdwXe>-x$PSqdT;(&@<V#U<=9*YdnQ!VADv&r2
z4-*SX)IaHu>zrP%uXP%Y>|p}?DnFYT^&e(eJ)8p(>Qb+bGr2z&0q^*69@{!`?)=H(
zbCbep-VkXTD&eoE2W&c9?qgD0JwxXc4Tw$n#)sbo9$MKWd<NPk;BKw-h^6@<PQLX!
zM@HGe_hc)3oT&7$MmUT!$NUgavto6HDh`FGhlnmpE)9;WnKCUFDE5<AJEf;^6)E1E
zVrfIk9e$hyv}8mC!aGzGmq_PaTH@4f`k^JFiFI=*rL6p+>UhX)<^Av#IZjfJ;6UAc
zhk$SrL%+W@mc<Wn^_6jI^2upT{Ead2gN=GGJF{9vww2%`c>7M5Top2Wlz2F{+r(a#
zKa3Hp9IMPm#JpCC=6L=aqq$q%MF)41tq7Rr^suv&({p{Po5sW+YOXPMJH!0Y8owHL
znhjFg+4=8$jYUS?_ILgUY>zfJtIf|sz@ybX;@I)axQIR`L=)fY<bys2b+H5@f}jxh
zen@pP^dphjX*lfOiNYEW5IA@nLm{0)%inbZ%O05~jYA3x9>aShq|)=OQ4=eJ&!KTz
zUQr^}#KX!zf3+lV3AG&fSqnzd13zo=UZjg)Xm)0Dc0l5A0H@7gWoHO9(lGLbi>oiR
z+-dufh_Us)`RcQNN|q=rUtz-7e$WdQT00c#&ftX77!mHx&CeZ3C<I|J`*gd7mu@x*
zS79H63Py|%vXrfLIngDzl&J3EhCf~R^NcqUE<b3)P#<5=VNZ4DZCEVlS=J*|D)1|<
z%8+F3KQNi&rag0o)39egbd{lI6~$l|fDT+Cu6ek9b*po=OS31dR~ywfzWz|NwiPii
zT#gar2_nC8be+c5Y0&c!yENh;s)TQ+dyEwMklIaN^L$CJ9)<l+Yu?x%kLj!Rt??Ac
zZ#4J`6SmOcr&?p8yvt$~2kq}I1rF=xX66T^0IzV0Ad;T|J=f-kD;PXoB1uppT%-eX
z&-yTHA6M&_I%|A%F&rx7LQD@(v~$(<Dp~SwtubEV_Zf+2)3U||5^`+l!8{GiJY#_-
zgQD0*X-9X{_v7Y}`zhpyNYOOWwV-y4uPXyyFHipXqZ#avurPIC`hdaNfw}(Bj^8f<
zvqp2Nx`l~2$InPFRT()Gf-#aMr=NAMu4~;Ms`^uy7sv(bx<fbl9qs$;?Uf!9bZqC3
zyHpAug_$_w@P&0su>TbOW>{XmOxWTsADnnz=MblVuD8<Is%}8+Xn|PR4GFr$IRx?U
ze?3;Y(U@c6f1WXC4hPnz$bEWt*3g||9+dbMYxr`g(`oSorDBI$Se^(I71E9m-O0}#
zQL9*{6v(6W0sM<k#C98$rIB%B>M(t7zE4GmW~L@*>Cnv7)O6q-sT|y9azrt7x>{eU
z*APuY?f7NEN2^UdIX{KLXbbj{-a1a2i!(>81UX3RD(gRTukZx39?b^7KX;+^*(P5)
zpEYyQTr)b*4q9hq0T7q|%xl==G&;Z$@_V5%9vjR*F%8Is-`qatg$Wd;wrer_G5!q$
zf+toR7mK7%ZbT%x@e1B|gT+Mr%0jDm6$7NLUcm-ePU>jM-|>)N+F_X*Ydxg*@g8hH
zQYG2;Y242P9z49xehWNSyuiAdD!=TSe)#hT+y56+Q$tC(p<jUO#IJTPm^!P)R-000
zrMk()Q~IWCwa3>tA=3W|&t3#X`%KZzPtE4;N#F1d(sC|neTZab3dbIo5z&>$dJKIy
zxxjx6Hl@{xe3W^C!0$04{Cogi{P16%*^RjoXD-TL)nI&@|87RgxDI!{Oex;L)P3ZQ
zCTAJ-6<&U7>V07RJFJ%f{M`rkRrgQtKQJ@Cyl?-$@!9$Pwejlg^vrm*I#pkqKd^sa
z^}zCHChuOzKL6ZU_Mdxt>ZYEazNx2YZtCgTn|gZgrk>t+Q%~=|si)^}>gfYF_4L$D
zGV|uRIdyaFoVq!FPTd?sr*4j;Q#Z%bshi{J)Xgz<>gKpQeREu$zB#T=-yBz`Z;q?e
UH^<fKo8#*A&2e=)arL?XAHTu%5&!@I

literal 0
HcmV?d00001

diff --git a/vlite/main.py b/vlite/main.py
index 70c8b60..25edbc8 100644
--- a/vlite/main.py
+++ b/vlite/main.py
@@ -64,7 +64,7 @@ def add(self, data, metadata=None, item_id=None, need_chunks=True, fast=True):
             all_metadata.extend([item_metadata] * len(chunks))
             all_ids.extend([item_id] * len(chunks))
 
-        encoded_data = self.model.embed(all_chunks, device=self.device)
+        encoded_data = self.model.encode_with_onnx(all_chunks)
         binary_encoded_data = self.model.quantize(encoded_data, precision="binary")
 
         for idx, (chunk, binary_vector, metadata) in enumerate(zip(all_chunks, binary_encoded_data, all_metadata)):
@@ -88,7 +88,7 @@ def retrieve(self, text=None, top_k=5, metadata=None, return_scores=False):
             print(f"Retrieving top {top_k} similar texts for query: {text}")
             
             # Embed and quantize the query text
-            query_vectors = self.model.embed(text, device=self.device)
+            query_vectors = self.model.encode_with_onnx([text])
             query_binary_vectors = self.model.quantize(query_vectors, precision="binary")
             
             # Perform search on the query binary vectors
diff --git a/vlite/model.py b/vlite/model.py
index 64970e9..413f3cd 100644
--- a/vlite/model.py
+++ b/vlite/model.py
@@ -4,10 +4,6 @@
 import numpy as np
 from typing import List
 from tokenizers import Tokenizer
-import numpy as np
-from typing import List
-from sentence_transformers import SentenceTransformer
-
 
 def normalize(v):
     if v.ndim == 1:
@@ -16,11 +12,17 @@ def normalize(v):
     norm[norm == 0] = 1e-12
     return v / norm
 
-
-
 class EmbeddingModel:
     def __init__(self, model_name="mixedbread-ai/mxbai-embed-large-v1"):
-        self.model = SentenceTransformer(model_name)
+        tokenizer_path = hf_hub_download(repo_id=model_name, filename="tokenizer.json")
+        model_path = hf_hub_download(repo_id=model_name, filename="onnx/model.onnx")
+
+        self.tokenizer = Tokenizer.from_file(tokenizer_path)
+        self.tokenizer.enable_truncation(max_length=512)
+        self.tokenizer.enable_padding(pad_id=0, pad_token="[PAD]", length=512)
+
+        self.model = ort.InferenceSession(model_path)
+
         self.model_metadata = {
             "bert.embedding_length": 512,
             "bert.context_length": 512
@@ -28,20 +30,70 @@ def __init__(self, model_name="mixedbread-ai/mxbai-embed-large-v1"):
         self.embedding_size = self.model_metadata.get("bert.embedding_length", 1024)
         self.context_length = self.model_metadata.get("bert.context_length", 512)
         self.embedding_dtype = "float32"
-    
+        
+
+    def encode_with_onnx(self, texts):
+            # Ensure all text items are strings
+            if not all(isinstance(text, str) for text in texts):
+                raise ValueError("All items in the 'texts' list should be strings.")
+
+            try:
+                # Tokenize texts and convert to the correct format
+                inputs = self.tokenizer.encode_batch(texts)
+                input_ids = np.array([x.ids for x in inputs], dtype=np.int64)
+                attention_mask = np.array([x.attention_mask for x in inputs], dtype=np.int64)
+                token_type_ids = np.zeros_like(input_ids, dtype=np.int64)  # Add token_type_ids input
+
+                ort_inputs = {
+                    self.model.get_inputs()[0].name: input_ids,
+                    self.model.get_inputs()[1].name: attention_mask,
+                    self.model.get_inputs()[2].name: token_type_ids  # Add token_type_ids input
+                }
+
+                ort_outs = self.model.run(None, ort_inputs)
+                embeddings = ort_outs[0]
+                return embeddings
+            except Exception as e:
+                print(f"Failed during ONNX encoding: {e}")
+                raise
+
+
+
+
+
     def embed(self, texts, max_seq_length=512, device="cpu", batch_size=32):
         if isinstance(texts, str):
             texts = [texts]  # Ensure texts is always a list
-        embeddings = self.model.encode(texts, device=device, batch_size=batch_size, normalize_embeddings=True)
-        return embeddings
 
+        all_embeddings = []
+        for i in range(0, len(texts), batch_size):
+            batch = texts[i:i + batch_size]
+            encoded = [self.tokenizer.encode(d) for d in batch]
+            input_ids = np.array([e.ids for e in encoded])
+            attention_mask = np.array([e.attention_mask for e in encoded])
+            token_type_ids = np.zeros_like(input_ids, dtype=np.int64)
+
+            onnx_input = {
+                "input_ids": input_ids,
+                "attention_mask": attention_mask,
+                "token_type_ids": token_type_ids,
+            }
+
+            model_output = self.model.run(None, onnx_input)
+            last_hidden_state = model_output[0]
+            input_mask_expanded = np.broadcast_to(np.expand_dims(attention_mask, -1), last_hidden_state.shape)
+            embeddings = np.sum(last_hidden_state * input_mask_expanded, 1) / np.clip(input_mask_expanded.sum(1), a_min=1e-9, a_max=None)
+            embeddings = normalize(embeddings).astype(np.float32)
+            all_embeddings.append(embeddings)
+
+        return np.concatenate(all_embeddings)
 
     def quantize(self, embeddings, precision="binary"):
         # first normalize_embeddings to unit length
         embeddings = normalize(embeddings)
         # slice to get MRL embeddings
         embeddings_slice = embeddings[..., :512]
-        
+
         if precision == "binary":
             return self._binary_quantize(embeddings_slice)
         else:
@@ -58,7 +110,7 @@ def search(self, query_embedding, embeddings, top_k):
         # Convert embeddings to a numpy array for efficient operations if not already.
         embeddings = np.array(embeddings)
         distances = np.array([self.hamming_distance(query_embedding, emb) for emb in embeddings])
-        
+
         # Find the indices of the top_k smallest distances
         top_k_indices = np.argsort(distances)[:top_k]
         return top_k_indices, distances[top_k_indices]
\ No newline at end of file