From b2088aec49f7a5d8c35990feac8c238369d93899 Mon Sep 17 00:00:00 2001 From: wangmeihua <13383952685@163.com> Date: Fri, 28 Nov 2025 16:20:20 +0800 Subject: [PATCH] rag --- models/embeddingmode.xlsx | Bin 15810 -> 15866 bytes rag/rag_operations.py | 316 ++++++++++++++++++++++++++++++++------ rag/ragapi.py | 246 +++++++++++++++++++++++++---- rag/service_opts.py | 2 +- wwwroot/test.ui | 5 + wwwroot/test_query.dspy | 7 +- 6 files changed, 488 insertions(+), 88 deletions(-) diff --git a/models/embeddingmode.xlsx b/models/embeddingmode.xlsx index 55a8a769113707df5d4783f488a170a9073bb456..af4ba5d97e3dacdfb1cc288564aa72fb701815b8 100644 GIT binary patch delta 9648 zcmZX4WmFzbwlxymoe(ey!B9Ft6;M8zY8KFaATSZ|1^FRwqB)K>uTx_rNaXmtlppY%`uvd z`KiJf8rdzB;?IQVf+bp+|KQ;Erq|p?f-m2dxoTC9&xh2ir(t00Br$RE)b5kfH>5^2%&#GCY)M6|0J;N`c!BPSu7WC##(Nm{^0u1TR^ z{g6yC6730r8Fc?hr0ff+Bn$Zg^Syy#!NL@FgxHIaKs4{oUdIiY`~hB>bvGaa5)8~A z4wQrN5n%b-?hP8UE*VUyogRg6!V5P#OpUu-grby#@}RWFIqN4bTH>{ISTPFYgqNpE zRHNWIHvXHy0{zpbO#$*6zOE^*Rn?NzVNrZfvS_R9@WlN!RN&41(INKREV&{a4mZ*z zr1}G&_a7T=IIkl2M4Jz=E%28LCnnp^G%GVeG?gMI8_a~naI}@yXhjvgw6xAA^Pfa< z60Cw`bVJz2B!+<`9G{Y>8>B{~4^{)EMkAnKZHC|%^45zOdu1s~h!>L;<91eUc-oD; z`x#9mFlFrOQ^Loa=412uuBS4&>h@rdBFxOiHj)N-!5D~}bqCY@i~H>1l>NF(N=M58 z_UjsAruMPg^YYj zlG==5Co4!QLVD!Y2#Jp^VW~`|pnEjf}qK zbeN56%DTZ1gNtS_IH0+I3Qs8HLDHsAEROH?1bYpUxSlZ<&13_wrA*-Fk z@ksn?MO4sRUck_W?l)bY1(XHy*h{Qq6p@>e8C~Y%p8KU>+7)2hvzJ+pqH!Bv(o#xB z1n`pAWni+t?1K?_kHdZVwFqS*WBh4DN}l+DkomLG0*020Pt!^rs2A=@m@8<8PnQbK z92pIlOEcs~6ugPvQpQw_3uhq`^fF1YX=7@0T+A_5*ty`_k-+j@#)4A zJl;rwGk)_M%zZ(9l_>vGUhxkn#!Yx66tlL~9soXsC6@2#F$T3&U&s&X3~9kU0A5AO ztB;W6_5uQKWGk+qlFlj|fuw<4E+e$sD1{xK?$6m2^%K&6_g}`ZKX|&{hs;J$HUbGC z@3_W<*nzSPOa~jBR252_E>kDC4#1a-*2JBakA9&OYvhxTUHkK-{N?xHgwhSmZxii; zfqN-|(UE;q>gibO3G<|#X#^${tfY@s9MrFuf1Yk2$n2$(&`4;LcRm+h?(h6*D@d{_ zLl{I1giTU#HkKoJ?kt3&jkq(Lpt;xsRFe9;HSB%$X(d=ZiZT5$Kd5$*xA*D?YjseB zj=LzVBo)Yy(9A(8%b`NPxb@hsOZtp&U^1!}gy{JQz+hJKOrbJn;$PBotE-It^i)gs zmptVfYY`;c*PhK60;n~l=Nn5DNt}>I_?>ztK(bAr`^Q}}8|^z97YQrw=#b$7%t>F1 z#Tbs-cGA!-@f{wgq8IT5=i)qjeib^!OP1)RL)_nIQ9hG8eiyO+vfBRnVqXz^xna$h zff{*ipE9h$phR7t7fM-YBeXQ5`9@eYEqA+#55AqzLb_c%dsGIJ*egW2D4RZbOA7KH z@aA-%W0-8I{jbI`sk3)C?$jHB9NbMz=m$Ce_$s^TEv)Bpce$^u1E3nDz=GufRcbJ} z?BYEpBum_+y_pL8Fmz$AUotIWNCbyxnWhf1?NW($Xv2)bFE%D&WYJT(i1T3w?G0)wN@W}lpD`65g_gK;7RAJ750{Y{mO;7c0B z!m9XZ;AgK9^xi?8=CHiBA^p!NncnvK=_=hk4*YaRE#wZA_Ky;!8N&AtW~b%E1u)iI zl)pTnUQlrn#kYE@_f^vUA!SZ|mnN+oL$~AV9{x{sVqMACOnLvi<=_8qJb>+QIEqo1 z{~L}v-osJA48Lz`e=PrC+Mq&VmNjo$A(5pvodiZeF?mKoS0}A;H8cK~p4nfQcO451 z1{)pNy}Vy%^^)NtUgABKSkkV(q{JHv5>+`w6d}>8(wM=)c|ppbX%Gq0*_If%A{+6U zQMrkHF)bBGwMrHhZeRpxsV>PurMb{eHaJHz;PUITA+VIjTuAVvva?Kd^FCZ=$+n_-PL5Zc*&(6C^$6d@xAAMdLtGaK0<$0<&>-0mq>i$bfNC5d z?;Sk8+dTadWjfC>sxgWqdklt^~X)dDS{a%$r>TrEJ9PBmT4iugIAsu`^LEVSpZRPm+N)EKhXB1d-+*4`%Ob zP#sKWV8=$Hl{6{q`faB(;O6&aD`q9>3CSLLFA0aJ*z%z#E-7UbLBSz8p?z#vw9wJ6d9elV z&(+~kamR4(rX*-V)Ki9`b2cNvPyxz`0=84=p*(IMS7F}#b)ueMg8w}A`r5 zgF0m=Etx?1n`aZA>zm!3j5Z2QJtc;ZZJRl2s!3J)PvD_dm7Yp>owFrm%^xSVQqzk& zJ)I{UJBrQfsuW8##QZH8&%3Px-mNN1G+&1o>sUQY7q@FgGb^pEBhj%`^X?JMC2Poq zx&K+3dkJAzbJw;sZ%M+j!tjVPHT2FqxvA%r@!Bp@GbDV#=IZrvx+AQYK<=!hl=j&# z^Nf$!8~_$ShV~Hg?BbkSW8xrS8o-P?AP@dJ5NK-Asv0El{2g>#T5g2pe9~cH&+3fB zz-dheGCvs3Vwv;gJ>rg{|5^<8<4B&mh^L?1q#?+I`SfbF`pe-~M#x zlvAk1#iiN#B3Q&1UC(!CP}2DLz44PiP$AO>1F(oPbNVm#8J4b|($^OXT$G?RR_fd; z6R0D_%88|_9pR1kRxPQ8H(sQ_n71PrC6q9lrw?avA#HP_88BDQW_r@FtQZUA6I6co zA%6?g&;IR{b-LuAI!cWPl-wAl_cfa}dz`3O!JssUj-)aS6*1_^EC)k;IEb|I2n2g(uDW}%`+%>ravjRc|*EM+GUlCzxOYhMipL(UCoPLkc;DrUt$gWcPbCo?YLYnkQCC6~*`qv)J!j#x z@hDFNr9wYG#&1z6;fGM~W``7CF3~UB|5~++F&JEkkMK$E`XqXwOKt6@=E;<$EdP(Z_E_dx|P!NUl7tR{G21xbWrT&`oW#b(0P#k#u zb^4d}?zAQ`-jK zOHJ;2N$a(f^D$D*dGRquM5P4(ub%T|!qvOc!4KdsCXk&nKOj5$uQ!5#dEE4!Tk+wM za(`6t8-3Paz^l^G*U2gX0(TsbmI2+X1I+WIZ@MYRO-KLEVggopG`)=RRc#V|V z`uWt@u7CISk#*(?=D+*CR43u(<9pz0f&~M^eP3*O+A(@KegQFIVE`KXj;l?((B+iQ7N2HMm(F>tcl2OYa;ndJh7LDe>@D5y(=9j9 z60vJxwVCi(Q6jY{^Oi$?AI621=^j}3MFm^devSVFHF35`Bv?*UVsJy)?K!M;6n`2+ zozhvg%fF`=X9fpFS^}tXb8@Nk8m7Q|sxXeS0xX16e=?&2*%4)L36awZNkgRu!|>tf z;;Dw4h5vXn%B9{$=_G9Yn|7~zkSOYL%NtsrKKEh;TlR{y1}*IsgjMFlip_Sm7U4)g7V*4>u6 zZz(^XsFPMlF208_yxM1@k5HWA!yb=`7B;rQXA4;P{6G$CkR8=RCEqq9>VGl z*T(z7J5)Qq&X>DKC;AVq5H-D^xaIKCFa{(NvD784IDSj@8^$h=@k?2-L~ZNnk;sax zX5Q~H%llg86bvBZ1pjWGb5PZ)^jaDz~=aaJT#Q?Je-azt!+bw0InQ$^J&l=GgQ97Q9nBWQjj;( z>Tw!$qMK+p98}YX_Cg~up%geYHUi9G<}}mteJmVWtq#P!5-<6 z)O&`*SiRR3>DEv$0?ok%G$GqLexfNVj0sJr)g{6rU*L8zD<XZK*Yc6VerJeZ|>MDBEMzzDfxkz!uI zgw>T~=XB)fs4dT5ew&)z8k?QGd+3HVsN6+^SNOqp`hf^+?D@lik~amTFn*$3%U%&A+DGA;|S2VsY)k~H>9 z&F06iBfV_x9davhmu|Qf~jAs8nx-6doZWNMY}{q$o#g3ZL~jrLZGO3b4qb{Z8#SvH4j=%s6Z%$MuC0&<#{lTcT>{xWr z>zNMQ`qAm>jqIy}{i0}jes)B^*n|~uwm7Yc57r!NM?9Akl`itS@L;_M;8jRxs5` z?gb&ZUB7KDU@;J!x;**ZywRQP`PgP1aknM$S?mJ?5r(2yViLa@7vH3##*#$8dvjn? zwq4d1A8UFSQQ!i>X0+I&+HeNj*6Ja})+XvJ8TdZ?Gx{r)O)K7K4jAE|YreGb_XVm{ za2TQ-FvT9|@-BM@ol<&k`m(0>aizC))jYW~g*h{ALU+~hSb=K#xKgW(dB?2?8Z22v z^Q26O*-+H`F_Yyb>GAThx^0_Fwb91d#4iMSPRKbk2GgeMH7` zHhJg${0rt$un0y%h>t3xlY!8QwQKSKxg;1Wmx#RP>5^kr{M@>DnpkQAhB>HX2=Q5E z>~~Ubdw*o@H_hsYz0bH(SVpDjHFU!XOtT>@1oTc2z7Q~zVrc;dkhQt;dLkB$=0H&W z_ooPKE^wL%&n>xz65aE6&{8;~SslB+WDGjHRftc`60P2g>kwSv4f60ixji%L3Pq`b;)9FG`kUPM)*x6^WlIf91>Z(J#&}(@B@vTfjNKi`Z1YzvdL%oYmc;2J4o4 zw>i0&;;zi32CI=%yo*v2T^49Gf`_7xcnmDMSWU(Dq86n};ij|uHO^LT&sN4^yA1(v z7c$^V08?tr!n@mP*NQE4diZvxUZqe~J%dHNW>N;15R%0CrSe95f6TlKu6WGvVe zgPi(Vtz#oX!WNx?|04m`52fx? z)iW4n5qpC2Ttr+m4MKpm=3rlhOO0wRva>Y?4Vj^ScVsK@l(fN-bD97{)^U(hPfeHd zpqjmw-kTS0eK%mAY-^s1-bP>97I8^oV5V`o1jXMz=WPY#HL!TNZS(m>8IDuOiKl-j zp!jCsc*p5ltGr2aBWmI<_fcTu%N88(^Y$u3E<@>g|5CP=jvBz7q1pV@W7OcMp&b`D z%URP-{)7d!ZY{ySp=M;v@UpS=BGsQlg@Kba^Qwcy;ojRom`L;aWyo{pG$cl`mh&Ja zc3@O;1AIg%n4DkO69z|&PGEOn1U;&b^WYX*4~OV$RANA* z6rr~o&7CB=#tHm~VMJ^o_!F2uZ^clI(iC4(zjJZ#otZJ8V>OxBkIW%sU06f@mKS$Q ztkBA3$ae}`_MuJ04v-Yvj4H2;buSd&5bz0yA|C>=u zAWC>zf9&;ppmSw0wf>>6Rw6C}p=OS5IoMkRGgM-MO@h^GZlwH8DWKPeB3VD_Dge2f zxKXEZ9VnEx6u4w`I_qdWZy9m>oi@>TEEQvT_^0jgXG(|PpXPHa4|>4J$3b*~BdXM- zg`mo|4F>6!w`n2EvoiFDbw(dA3H8Wd{NpMvRrIKiMduNLxg{i-#0kYNAQXoI6c zn?`$nyKrn*SHM)~YyNkf_L?65BaF?~{NfNP%!4w;D|4~8=N&lSZj(X>;}TWQl@K|Z zM9L)C9}d``f#170AI_GlrZ{F(_XK~?-5T`!{c^Cf-1+ly6tnk;qsb?GEU5eh)%?qk z7Jg63mNoa>G%Iouf~iYi3b1<7#AB?c^`>Qv{5r!9@5^ujx8*ySX(?Re5#l8)47i-% zYhYSmYP!wHWv@`~K9yQy9}9}qfH||5vas8QHaF2PT>?iS^xY4#_=F$3703}R|!+zccFSj$ao(u>@lm$RreNML2t%&fJAtxT<9uZ%BBDIdqkS@Id7 zu1ACGrUwt~vWa7^4AR;5#s`8OplM_+M#?p?u{q3i!ZApz>$x)J5ZD~mJfd9e_aM~B zA>^+dvjK)Iw61+^kCNHsnusYT#MYoy-#!jetq!Opqsj`S!fF|>%jA&lblMAZi3uF+ z`ZlJogvqB|&O>-Km~FE^Llyju9xLIRk}e+ZM$m4W^>E|jDK&``OB9J7PBG+**eO2s zbi4d!nAyn77$x?;R0lbuMU zz0gZ)l+{``-X)VE#4GD(mv>}6T?HA( z#5kb~&)J(H3;Yy%HM5aS1?EdY`daJ)f)Vll8cf(8=SA#r^(X>_7K*Kd$LU?bVIu(- zofzf})QYZHsX%dpOAe zUYqyMDLb%!TjKX0XGF;*)*S)^Zg@&*{T=AEwUqfw?@bRf;)|%Hk#;0vA3lnPQymz;N6sl5vbhoHF^m78I2KJjH%JXxw zBfG96-+HX34-_TZGyiz;z=Y2I83+R&g2?uyq)Hi5Ce?E?Uu`EfG8fwP;<>Jr)V8Xjs*7rA$+n@G4Cg2~| z-xBMgM^^?UKUVO=rg;0?qAap>HNfoty;-LjYVLeJY!Wh|U&<5u^++ENoBX7MIgBnQ zSo8yCyS5%1sy1;GBEqQ!`?jY33q69TdUnN3{=QeM?(eIy>rv4?ws)W_FN8b_~kBf9lfP`2A*ABfkONKNS18Zj}otEtmpbpqMGp5^`Mu@x<`_Zm&w zLBC6JrM`6F%J*y-%!`k!)8h0ivm^rPSEQvg$d>@|Vl%Oyox4;szhGS?K%AJB#M64p=$hTR)>cyc`rOfhb}*^5D62%G)3g`Np5Q<{A2%8AW>Vxm-E9-{j9J_)G?=py-jMIEw#v zQiW4=E)F66ISWImC$azJWyMgU%ggcQdF~Phh&qyL`Z!bL2rp;F^LlZ!lP@TS=^6n3 zO==x@y)jjC@I5b~icxrWS*^t+lFLN@tjPl)Ohp*YNSWwLeXug&{-zK~Mryq*Li$xz zio6#xSv`TGC!ge7ILck}hFSHbp#UC#_>AykNBolUsvK!pe5OCPEE*cPixx{lonV)M z4?ZY7_3Jr|wA3Kc=BKygk)!Q%36!%q!56+lENAdJK6D-EM`|4&gkk+E*|nR8oT6R8 z=PjmxhMHT9bYT8yBdt@|Y+Ys9LYI575QBHR8JoPt+fq>+cL%ljk3 zkh|l#ReUMEL+a3&$2y^y(<0G3dRM}S>7~DA9AQmLh*gq^SAL*%wxEwl4MADeho;Sp z3D;g;qA>>jh2Pmk0Z&KrAsq-(Q7(O9)g^p`&q|q>Q^YmC0`S z?d)GUkG707KR&sO{<+*xXy|OS{Pn3}N5Q_2N_#)X+so1;->%->-DARAnX%Cy;v324 z?oXb@=rX_SS}n%;6%X!h3iyd-Jgy6KIYBlKwO8z}MWq=y)tvfih~kOYA_+|B%X^WG zI)JtO19%9Cj!Ou<1!T{~3ce2d&IN+{=QI(Bn_CT>92Cnf1P*|vxS1fJ*g@bt>=1P9 zAXIjIkQxu_r+=?M;lRND9*X&UfrAmh-;Y25&*xA7nNEX&A^ip4uYciPpdKC`@Dk7q z4+r=(h?Q3eLX{Ha#>)$y3##Geg&6+;>gK}&LGoeY|L0K6zjDF<(ddQwha~+@`u9FQ z^#3IVApAp$(7Y|E8n=^9qVs k{}7G&u|EBijC)^;lfQrH-<;Arb z_aE8WnP;BK+AEVxCi(5t<5SfR4a^0#Vn{CU^aIjF;kF#zBoK_}Qa*&eFcA zzDX^QD#3Wuy;L$Skc%x2Z9P4taOQfYUeRlSr#o2x<3rvn+1FLpT88f|asi86j&t)= zg0H1r2$}gTxDHX*6ZL3jOxM{>k-n1m&Z~s|8JnqOj?Mh%WLwEGcudZQu7fU^Xse&Y zC*&CD%gF|1T3CbK_&z!+XZ{Ss!!L;yv!VSlTzs{~jRk_cG7hzAC` zWgX?K>;E9hMUK)@hJPb3YHWd4W33QPvo3SXscSsRF_cEONmxE_66KE)yO|#>Ng{S} z?(#T|m(r*{OV}u}Ms)d6FZlHU^)5FLtpr>stTKHVBYePtk%n!fOQO%S2t{{NM$Wvc2~v|R?UUk_Dy-O9QkX; z@D(QA{@=<8g&cy~CptMU8)yP{k;u`;Wt$_j7J6g(x77TAX$+kfj@whvE(~J^@iwho z4h+gdm}kvTOtp!+`SP%qu6bI?^`3c$&=ypk8_Ob88H`HO-Ju-=S%D9xm1Gu6On%|* zs!&(5S)cT>pkl6(lW#(mT^9zZden9|A#$M$-D5nPN@-=GvK1I)Zuu?Vq=$sX$DI6b zri;<`MA)i;iK6OVu=dhaY9^TrnXxm5xAEW`jv4IAOYBP7hOdIu0W7k{c?(O)G^F2M zmdoAe7#1b758ocb!F)J|?+*w4;3`LPTV2-$&ZG?}sDaVjcUH8$R(}>R5 zuAk|WPA46%+He0|n_*`<>$aJ9?qa!b(#UfG!B~IP?UQVbYV>XeWs zd6wnJ$M4-D43Oigv3hzEhPBZ~iC);{^6T>~iJY&vnBM+O!KVaj?}B_JKWOE~#cOU< zWyN@Vas>pNpUci@a`zO(=;IVBOX^mYBETp5Kkwr`!~AhvyqAd)2`WtGyBO*4Rz>g2 zRM`WL#2^N^{`hUbmXT8$JDy)w-Sd{7AK5zxFoJjqjZwYAu_h$X$0HJ18_lu-8RGv~ z0SOMZDEZ9^fj@6|J{7njPdoWpO<+A5zgIKtLMlaA9Xb-ocT5UAdpthDP&>$JV7($q z+<9GivHJi?D@(G?V;Vq%36|M!;w)6|Rhx~N2c!^dEYQsrif!MU6ffjB;!}s*7ZI4^ zNQ0~B+PcYKX;pT|SPV0?#?_VTnF)u#4@0ePLz z<)jz|m{~X66&Y!#c#OL8f~MN)4_hI{u7Ggw{Ss&yF!wbg+Tud)>Zsugjl4E2V*wHDdZ`VK%r7>?f}hOmflsz+Hb4f2Vk| z7e1?D4O`z0FBOm$`pbjdVw)g%#>mktS@FAzD~|%5BxkU#r2Oj4P8@IQ&S%JDC&W{>_0bwb!u2N9-Syq! zNU!b-D8~mH%i8+NQJ9>m5WJR0 zNxG#%4KpN&Dt<7qwQ_0snm3I4@ekv|otdHD+76$~P|k$r7?R>xjxkh?${%BG84 z8y(0XGjb+-V{d+(yOoboOIA$IK}#+M#jYDP;@DkB@d%?ezO>?X#p*CS<4{^Y%=mK9#xJR25d~Nkv|lt0c@0uA(HN-TkOZ>8a07VRj?sp zZ-zKR8Ij4ZrL7ec2W0r*29Yfs##c?FEvdi0(m$_|#*LV{4V89J!l*E7kQ{xl+T1Bd zl&QW4qR-P9`3_^@tKiinmU4v`%JG}op?tEoj=r$h^4c`_Bk!ru#BDfd)8XbhZAzBL z29Z%ZmwAr{x+og5h?nJxMhpGAk~>Su9dUI=l84xq>5f$CQ_v-eS zk1-%jnKC}O*nm9YT~&XGf*r<$eNe%zwN8d|u}`6TCK}!)IImU0k-^$$<z!z40i%*7wtCw^gGa`#>}~jzwX$tIwvg`L(iq zv|G3Kqp8h`HTlY_-iGYW=2$JuCLR?*saEf3F@PE3lf34(K6eUSa<5y6t@xF5MVawm z^D?)N(mN8+cYLW^%_!r7Dt>YK4=Xye>~0lHYhyUGhP1^@=UvH zle;^MrRiO%61fVb?47l^%2IL>DDIqDKIkomhn%MNeg_t4Ao#8VSH(%7Ke_6Qn_-$U zsu-QMM_y&VNbdT}A&G37w!`By<+C7?F3MUEO(%woKiL=tG=1M=-F_phF7yCAUm zrlg(6G+|2o14u%Z_8GS2LEY{`%~SM&jiKWP)vTg zx}fqh{zGb<9}OCQq01P&r1EP6x*b*@G|bsbJ+7AQ=AlRL@!kCz`!BNVi?$zcj9vfj5dYBgAYWGdevo-3#&13z(xVB4$c@$CgFoRT0EvHts?f975K+p_q-dIaWju8r(|I}v_4?DD zl@Q4)l;mZVgX`muXz)A13hj*PjsBjHo27}p>jLImWDC z<5%W^49@ew{*cZ%+Carj2EOfrMZMPa-4O%5OJ*?YRNwmqO<{^-Cb&iZ=GWTnzZx?f zosi)pXe^ZnKVfp!4q+!H-Mn^_6{vf~OId2r);KAO&3*)yaeDpye1j)}OM z_46($NG2+iEaMJ$e^UQ8PNr{>ot|>K^i1u+El>ncTIvn_;u0UAU*a#5&&NJEDe~L_ z0V8hb$4r~$T({kht0%~sOaET^xZ#W7F@(>1G}AxuLsqK_rK1NKg%rSwz1uQU{m_ct z=2C?YbEU4ap5N^qkcn&5*2@@2(~;G1xjZ10Q~5({P4D{}!A)XQRsRxBulUu}sEIYN z4N~dAlq=zK%X(!0)=AoZHGgM=?10&{bz;_CK1CU^C?D?!tItdZYtj87PG72N5rV9V z^p%Kvj!djwGLvaDX$b&>Ts@KFwC#hm+8o6y+YX!WClb|C;N%b^mG{-QRdfj$cAqV` zS;{uv`KgjL>hdhrYI>M)a2M<9`mMctUoB`NiL1B2fTPSvK)`naM^K93L$Ex7vl-6Z zb!;10ItuIQfzdEGC1p#5;Y`vQXLW4q_pwDdrk4)~(|J%ilFSL{n=*F!usZY?_xZ}i`jPf0~yG=zjnGW^IX@ie$Hy@ zNh<(GF9H#T4gEmqR_Ucr@N7`yesCYkGEx^P$(KbRL;@Jxy3ZM%oX*4N4C%#FW)Qs) zTb&qw!w~tq+CSsL{flzEYaV62cG)-Hp8~W zF%)&wSt@`p|8{Tc9f@cBk$AcWFVp@Y?HRGM!P!Rn&3=~b^JI-` zN283&SAX;&T%zhu-RIXVtU4HxN2ycECj+Coh3^hC5>2qM09pI^YMi#7 z*&>wb*}kF~zPr-|Ss$~N(+L;I;{-_cy*Z~yZ>T!Nyn}3PTO?Jeu+t53HRHMC&XGOr z#@g|D;B>+SN06k&{+58wHZpxYW%~& z-^+&Zti`qjQdTF-*|b6g`|Fe9<=$rdgF%S!B*%C+uQVOw=~3oPe||qLj3lV;Qemn_ zawi(aC`S4a@uepc4?YtU|6V=>kGN88jh6Ope7o@6C8Jo5c08(kF|g((RZ)_yVd&Is zl~0el`xJ*$siqczBR!oUY4n1?W03tr6V@WA9{(}n7!UJ>qo6-u~^smac3du_0<=HQ0$jyaFs|07n zRvVx}F6O6C09pPrt@xcJB@Z2d$JetNHJh z#nL}mcCcf(o!vrE08x7Py@6m@56iF7?@Vy(ubIWG_phG!u5kZM11K-)2dKa42Pmjm zER5PgQZ@u+q_$Ij%9pqq4*X0-28^q$ieYu=90^Ii>4JAw&yOgS=hr;Kt@~CY@ugi} zd+B(twa{4+5R+3T!I;Wtz?Y`w4vV8CRtrGdfUm|7xIiHj!qaL+m7Vmx93N}s0-G0t z3}16Z`^gBW@~Gp=QwJ39E(a6T3|v~N4u1Ozw-dfHbU{Bs{3V%aw)y^s6Zm=o*QcJy z6Oo=YQ#f2k<$;}cu89jUKkyso-9a`Olm~Dv8M8UHEv(jk)|- zBziCWt=E#HFV#`@xY2V`Y1WBk;Utc{vz$oWQ~H58Rl^@_Bf6+AC!X-jU)S~QW&I|a zGRvJm2Sw=GhL*ETmyX2t0f{Wh65#Ef#kvT0#Cx#RV7W*CnpDy zLwT+m&~QgGt-*vokUNDEIk-MkWeX}2mwXZ)U9F?TF0MxVKDhwV#0xR1G#OHSOZ?cM z61J`sLMs@#eR=}uOx_)wT)Y8?QH}{;qaxJlx6B#rRO8iT5l1J7iNq`|Hp1H}d-*l# zS5Z1zq_8nCN)6(=0k1(x+fC!_R_KO}uPpq*QA(ua@)CUXRj6?3#9Gqz)Fr>Vf2Z(o zV7JykRc3qZ&2xzu{&sAv;9SKwo?F`S>=aS zaRz-<^Haqa60u90Q0o#|N_M@dzrP!KE8~mHG@7fUJq3}`1a=(Y_x5|vM8?UYF(n0A z8OtBV51md!kB#WHRBod$YG%=ov&@~wP|3c`N7tZ^%yD59d`^Hf5-T;m7t66~PM~-E z@mv1r+ats*{OozHHRp?Q(MmJFO~kxHx+W_}8BcGd@-hmWm&=v-rMfA^S?zelS>jh* z^tS=y`~iw~QW`~CuFFL`m{$I?yXYY_i%v# zx3hY`N#zQ&kEwl`QFFJDHqo1O;NH^7ScI5=#9@?TC3981{#Q+juH)kI@~OY#?2u8Q3hKYw|KT{WG5q8uKlxomG?qL zV|~S47S;_5Eun(XH7L=xFA2|Mf1KQKPBVmP{n|TwfMwTUB*E!jvwBBCn|Kt^6gt^) z2aW}`RS)jGz$}u3M!Jmo$P?% zPIW3L_-2y#MDv(`8Ta|^Ia*tNe-s#r?>TUExE4u5)T@+;A5JJA-xeByM8S?rQDYx@ zN9{8~XOrbBRW@Y=^k))gJstJeB3=~-MTrK>^dK=IG|QyP-TvqT5W3>N(9-20#XUT> zC)%c3{Ej~GC`PxNVxKY@=Qwc~{UH21pT?xPI;HpaLk>IDr&uV=&)nlm;*8*LGu25{ zRnSG&YTCs`L-fv3AJBro)H!0d;)+w-4<;zp)RjS&24>V2kb-phhTqY}68SMC>(Hi$ zaMzn$!r8>%jO4$v0$5+WyQ%4MFj=-Y)K$Hjl}N`5NZ8#r*Ut7T@6a%|rUqqGVOz%m zU)&GVZ_6iIbdJc(?Z3OB1(DY`ELDwTFW5`c>RK};GnHu@by%1-5uo^{y!{da56M8w z0ow|x2=f-t$u;6v*gN#_GgSF52IJBScX;_CWoUy!-SdYNfF(Bu)wBByGj#9X6y(rM zvzu^&dxtfx&{t9>;B5`B_6>X9(o}6W9b%b0iP8(>+bYO)88zyF2G5EG>nrLS#&k*w znUL5A#vvzd;_=iCdut2Zm7lwdmvIyXxN+924mbm}E5xI=jcMcf!!yda29afUT0a|! z^&l%1%Ns_K0JE5k--jBH#Gpzr zPBks8g<&tR+P@RgrPS_`RxL-9^!9MDo2~v0jYi|=^Ze-nVdE8%wzoKxVu?YTNNep` zGZ`X=BDVCEBE6{v5kahhtuxe@u4$zYR(PJ<-uj<~@my8?5|6~ResL1Y|O@5tBQ(~A4)4kt*ohb*#S$3I+!OooXP+RTYLbEDnfc`NQbCqC7ixyIUI@e zA zGaYNsZ-se!W}p*_>;xvz)j2yhqeo-KuXjm*P0J81^+@r)jcMoqfo0j=E2-_hG`wIr zSbeh|u-mhMHi~(|y1X7VnPhwA81e>Wf1Ko zIP7LKx@qq^LU30kC~l}MV_cVZW6%saRq(Eg*w$>RJXYtQ&o=014Jj$A?s1otFtfOL zKDDm0x}#fH9a@cOJ{kz&4voE}oh77|O77c1#)gvA&9?4)I0lom1S?i0kLG28iNmDQb$yiLe&2 zz3rY7+fmG>wKh~^ap?=;P|jK5Ur;}w##2!Is+jWzq^+1^0#Z;Uz4*RBo=?k=2=B0i zl$5V;48!upz{k07yD7V)PoJRi+Qjfr@lWH6q$;t$T!jDoOCdL^;9`dv;aSjBk7%d2 zwh}~*0T6@gUa&8^`rpI%jeQh6Z8BNr@_4)nKdZ77O1e!}o%qak9=!KMc$!M2ezI?| z6-xgVK;r%zWqi`Mhw33@j&Ow|#sxFLkR_*mqNu;{imIR~uCjwgjMd0Gj4262 z@Zy?>+V5fk1>z;Mq3(;Wp8MIeTb~~rddLHa4A&iP;DeVY=&F;jp+_$7;*C$%p+t{6 zsu7XbiBi8|RPL5C%mb90uzm!i27(mfQ0_j(epqk(eT}``kjB6=?FtPK<8X+>=ia|7 z^@@VXrg-O@>Lht*H1XKRkcWwoXenkDvSaAh(>iGKG#a?Lyu<`ZV#Ot;%un3#r=x%x zwTb0Uis-Bk66~11vE8YUw}rcH4?AvHoOeFV!Je_RQ=pkbn*q5-ig*A|9@d>&;dj(5 zH?0QdbWG1=!8FLy()!8m?4kQ|xROHJbqFhFwn{ZysoC`5&dTcBy9`lbzq{>wtJy1a z=Yhk+h9{d8F!1>6;YW}hCDCga7+v6dOyBXz@VkM|JCmRH;9*(d%fZkP{B(7+jFOGu z-R+MDu&5NC`yE`Y+&XZ5tD)xTx?fEn$!hDoQi4aJl#TLYpSqM( z7h1;GA6JuXsuTk>RCMkwZ+-NuhrfNK1H8FAo?9>4TfxztUu~of4`#$Th9;+G1XNDF zjaCX@KMSX<-8s=jsqji!`^U#*e;m$xjan^S8{Scfi)fdRx z-*Y?4Hu2;FVt?cR^ZWrEJ>d`B5rwh%_KPFc)L|d*I?>yW81hnSn_#gax5ZA#^kIIj zp|!U?w3kxS%1@3<#%)wIg432pDU+(~vp+CIKt#LOoa1i5ZSPrDN?>9*#d)c8VUOgV z>tTI%qc#4Fi`gV@LfLPXtHV!%U3+3ik@5lvA1arNO*_D!kddl*1xQ!)F}YZlhbMOJCEtf)MNI{2C{bLZ@uszeb?nOF&0l`|9X}kol<&uC5|Vu=UhQY& zZF{LPV)ZdP?~q#TSi66S5gvjfM^YW2-{J>U?xc}|khx-PQ8UafJMW^%zuelM+ zF}7;@l$fE@U9}ZFPy=_|f%*{=Bu**-7$vjj`i18kJ;+ zUaLz!5$K-PzFE(<#%ZyqFVWH-?G@RrcNqmS2_q2n&aU_%xL8%+8hOt0S%X@k_cpD? z;2AN7xOSE+1`bA<7o5Y35BBC~gbUz;f mj5t6GIn(m-)f)fA0-}T?8~?sA0ik zu=wCL0Y*fBxWDl5(%piyXGtLbb5afZm;s!elGL!sg$BGOE_2)$W&jSrb`BIjE z8wFpJ{BLCbe{yT^ouDAJ9hh5)7dj4XCnN^*KnpGx5`xYG9|{S!5UpfGT?FcazjoqhhV{1e7s5_A!K(y9NPTm3KS z8S4c^4AvFFBK_~=_W$1V<`MsCjJk86N 0: + # 或者保留更多有用字符 + text = re.sub(r'[^\u4e00-\u9fa5a-zA-Z0-9\s.;,\n/]', '', text) + else: + error(f"文件 {realpath} 无法提取任何文本内容") + text = "" # 确保为空字符串 + timings["load_file"] = time.time() - start_load debug(f"加载文件耗时: {timings['load_file']:.2f} 秒, 文本长度: {len(text)}") - if not text or not text.strip(): - raise ValueError(f"文件 {realpath} 加载为空") + # # 加载文件内容 + # text = fileloader(realpath) + # debug(f"pdf原生提取结果是：{text}") + # if len(text) == 0: + # debug(f"pdf原生提取失败，尝试扫描件提取") + # text = self.pdf_to_text(realpath) + # debug(f"pdf扫描件抽取的文本内容是：{text}") + # text = re.sub(r'[^\u4e00-\u9fa5a-zA-Z0-9\s.;,\n/]', '', text) + # timings["load_file"] = time.time() - start_load + # debug(f"加载文件耗时: {timings['load_file']:.2f} 秒, 文本长度: {len(text)}") + # + # if not text or not text.strip(): + # raise ValueError(f"文件 {realpath} 加载为空") # 分片处理 document = Document(page_content=text) @@ -67,6 +96,40 @@ class RagOperations: return chunks + def pdf_to_text( + self, + pdf_path: str, + output_txt: Optional[str] = None, + dpi: int = 300, + lang: str = 'chi_sim+chi_tra+eng' + ) -> str: + """ + 将扫描版 PDF 转为文字（你原来的代码，一行调用版） + + 参数: + pdf_path: PDF 文件路径（字符串） + output_txt: 如果提供，会自动保存到这个 txt 文件（可选） + dpi: 图片分辨率，默认 300（越高越清晰） + lang: 语言包，默认中文简体+繁体+英文 + + 返回: + 提取出的完整文字（字符串） + """ + # PDF 转图片 + images = convert_from_path(pdf_path, dpi=dpi) + + # OCR 识别 + text = '' + for img in images: + text += pytesseract.image_to_string(img, lang=lang) + '\n' + + # 可选：自动保存到文件 + if output_txt: + with open(output_txt, 'w', encoding='utf-8') as f: + f.write(text) + + return text + async def generate_embeddings(self, request, chunks: List[Document], service_params: Dict, userid: str, timings: Dict, transaction_mgr: TransactionManager = None) -> List[List[float]]: @@ -149,6 +212,14 @@ class RagOperations: return result + # async def force_l2_normalize(self, vector: List[float]) -> List[float]: + # """万无一失的 L2 归一化""" + # arr = np.array(vector, dtype=np.float32) + # norm = np.linalg.norm(arr) + # if norm == 0: + # return vector # 全零向量无法归一化 + # return (arr / norm).tolist() + # 统一插入向量库 async def insert_all_vectors( self, @@ -200,7 +271,10 @@ class RagOperations: # 遍历 multi_results for raw_key, info in multi_results.items(): typ = info["type"] - + # vector = info["vector"] + # debug(f"从后端传回来的向量数据是：{vector}") + # emb = await self.force_l2_normalize(info["vector"]) + # debug(f"归一化后的向量数据是：{emb}") # --- 文本 --- if typ == "text": # raw_key 就是原文 @@ -253,7 +327,7 @@ class RagOperations: # "upload_time": upload_time, # "file_type": "face", # }) - # continue + continue # --- 视频 --- if typ == "video": @@ -289,7 +363,7 @@ class RagOperations: # "upload_time": upload_time, # "file_type": "face", # }) - # continue + continue # --- 音频 --- if typ == "audio": @@ -776,25 +850,150 @@ class RagOperations: debug(f"三元组匹配总耗时: {timings['triplet_matching']:.3f} 秒") return all_triplets - async def generate_query_vector(self, request, text: str, service_params: Dict, - userid: str, timings: Dict) -> List[float]: - """生成查询向量""" - debug(f"生成查询向量: {text[:200]}...") + async def generate_query_vector( + self, + request, + text: str, + service_params: Dict, + userid: str, + timings: Dict, + embedding_mode: int = 0 + ) -> List[float]: + """生成查询向量（支持文本/多模态）""" + debug(f"生成查询向量: mode={embedding_mode}, text='{text[:100]}...'") start_vector = time.time() - query_vector = await self.api_service.get_embeddings( - request=request, - texts=[text], - upappid=service_params['embedding'], - apiname="BAAI/bge-m3", - user=userid - ) - if not query_vector or not all(len(vec) == 1024 for vec in query_vector): - raise ValueError("查询向量必须是长度为 1024 的浮点数列表") - query_vector = query_vector[0] + + if embedding_mode == 0: + # === 模式 0：纯文本嵌入（BAAI/bge-m3）=== + debug("使用 BAAI/bge-m3 文本嵌入") + vectors = await self.api_service.get_embeddings( + request=request, + texts=[text], + upappid=service_params['embedding'], + apiname="BAAI/bge-m3", + user=userid + ) + if not vectors or not isinstance(vectors, list) or len(vectors) == 0: + raise ValueError("bge-m3 返回空结果") + query_vector = vectors[0] + if len(query_vector) != 1024: + raise ValueError(f"bge-m3 返回向量维度错误: {len(query_vector)}") + + elif embedding_mode == 1: + # === 模式 1：多模态嵌入（black/clip）=== + debug("使用 black/clip 多模态嵌入") + inputs = [{"type": "text", "content": text}] + + result = await self.api_service.get_multi_embeddings( + request=request, + inputs=inputs, + upappid=service_params['embedding'], + apiname="black/clip", + user=userid + ) + + query_vector = None + for key, info in result.items(): + if info.get("type") == "error": + debug(f"CLIP 返回错误跳过: {info['error']}") + continue + if "vector" in info and isinstance(info["vector"], list) and len(info["vector"]) == 1024: + query_vector = info["vector"] + debug(f"成功获取 CLIP 向量（来自 {info['type']}）") + break + + if query_vector is None: + raise ValueError("black/clip 未返回任何有效 1024 维向量") + + else: + raise ValueError(f"不支持的 embedding_mode: {embedding_mode}") + + # 最终统一校验 + if not isinstance(query_vector, list) or len(query_vector) != 1024: + raise ValueError(f"查询向量必须是长度为 1024 的浮点数列表，实际: {len(query_vector)}") + timings["vector_generation"] = time.time() - start_vector - debug(f"生成查询向量耗时: {timings['vector_generation']:.3f} 秒") + debug(f"生成查询向量成功，耗时: {timings['vector_generation']:.3f} 秒，模式: {embedding_mode}") return query_vector + async def generate_image_vector( + self, + request, + img_path: str, + service_params: Dict, + userid: str, + timings: Dict, + embedding_mode: int = 0 + ) -> List[float]: + """生成查询向量（支持文本/多模态）""" + debug(f"生成查询向量: mode={embedding_mode}, image={img_path}") + start_vector = time.time() + + if embedding_mode == 0: + raise ValueError(f"纯文本没有这个功能，请重新选择服务") + + elif embedding_mode == 1: + # === 模式 1：多模态嵌入（black/clip）=== + debug("使用 black/clip 多模态嵌入") + inputs = [] + try: + ext = Path(img_path).suffix.lower() + if ext not in {".png", ".jpg", ".jpeg", ".webp", ".bmp"}: + ext = ".jpg" + + mime_map = { + ".png": "image/png", + ".jpg": "image/jpeg", + ".jpeg": "image/jpeg", + ".webp": "image/webp", + ".bmp": "image/bmp" + } + mime_type = mime_map.get(ext, "image/jpeg") + with open(img_path, "rb") as f: + b64 = base64.b64encode(f.read()).decode("utf-8") + data_uri = f"data:{mime_type};base64,{b64}" + + inputs.append({ + "type": "image", + "data": data_uri + }) + debug(f"已添加图像（{mime_type}, {len(b64) / 1024:.1f}KB）: {Path(img_path).name}") + + except Exception as e: + debug(f"图像处理失败，跳过: {img_path} → {e}") + + result = await self.api_service.get_multi_embeddings( + request=request, + inputs=inputs, + upappid=service_params['embedding'], + apiname="black/clip", + user=userid + ) + + image_vector = None + for key, info in result.items(): + if info.get("type") == "error": + debug(f"CLIP 返回错误跳过: {info['error']}") + continue + if "vector" in info and isinstance(info["vector"], list) and len(info["vector"]) == 1024: + image_vector = info["vector"] + debug(f"成功获取 CLIP 向量（来自 {info['type']}）") + break + + if image_vector is None: + raise ValueError("black/clip 未返回任何有效 1024 维向量") + + else: + raise ValueError(f"不支持的 embedding_mode: {embedding_mode}") + + # 最终统一校验 + if not isinstance(image_vector, list) or len(image_vector) != 1024: + raise ValueError(f"查询向量必须是长度为 1024 的浮点数列表，实际: {len(image_vector)}") + + timings["vector_generation"] = time.time() - start_vector + debug(f"生成查询向量成功，耗时: {timings['vector_generation']:.3f} 秒，模式: {embedding_mode}") + return image_vector + async def vector_search(self, request, query_vector: List[float], orgid: str, fiids: List[str], limit: int, service_params: Dict, userid: str, timings: Dict) -> List[Dict]: @@ -866,34 +1065,49 @@ class RagOperations: return unique_triples def format_search_results(self, results: List[Dict], limit: int) -> List[Dict]: - """格式化搜索结果为统一格式""" - formatted_results = [] - # for res in results[:limit]: - # score = res.get('rerank_score', res.get('distance', 0)) - # - # content = res.get('text', '') - # title = res.get('metadata', {}).get('filename', 'Untitled') - # document_id = res.get('metadata', {}).get('document_id', '') - # - # formatted_results.append({ - # "content": content, - # "title": title, - # "metadata": {"document_id": document_id, "score": score}, - # }) - #得分归一化 + formatted = [] for res in results[:limit]: - rerank_score = res.get('rerank_score', 0) - score = 1 / (1 + math.exp(-rerank_score)) if rerank_score is not None else 1 - res.get('distance', 0) - score = max(0.0, min(1.0, score)) - - content = res.get('text', '') - title = res.get('metadata', {}).get('filename', 'Untitled') - document_id = res.get('metadata', {}).get('document_id', '') - - formatted_results.append({ - "content": content, - "title": title, - "metadata": {"document_id": document_id, "score": score}, + # # 优先 rerank，其次用向量相似度（直接用，不要反） + # if res.get('rerank_score') is not None: + # score = res.get('rerank_score') + # else: + # score = res.get('distance', 0.0) + distance = res.get('distance', 0.0) + rerank_score = res.get('rerank_score', 0.0) + formatted.append({ + "content": res.get('text', ''), + "title": res.get('metadata', {}).get('filename', 'Untitled'), + "metadata": { + "document_id": res.get('metadata', {}).get('document_id', ''), + "distance": distance, + "rerank_score": rerank_score, + } }) + return formatted - return formatted_results \ No newline at end of file + # async def save_uploaded_photo(self, image_file: FileStorage, orgid: str) -> str: + # """ + # 把前端上传的图片保存到 /home/wangmeihua/kyrag/data/photo 目录下 + # 返回保存后的绝对路径（字符串），供 generate_img_vector 使用 + # """ + # if not image_file or not hasattr(image_file, "filename"): + # raise ValueError("无效的图片上传对象") + # + # # 为了安全，按 orgid 分目录存放（避免不同公司文件混在一起） + # org_dir = UPLOAD_PHOTO_DIR / orgid + # org_dir.mkdir(parents=True, exist_ok=True) + # + # # 生成唯一文件名，保留原始后缀 + # suffix = Path(image_file.filename).suffix.lower() + # if not suffix or suffix not in {".jpg", ".jpeg", ".png", ".webp", ".bmp", ".gif"}: + # suffix = ".jpg" + # + # unique_name = f"{uuid.uuid4().hex}{suffix}" + # save_path = org_dir / unique_name + # + # # 真正落盘 + # image_file.save(str(save_path)) + # debug(f"图片已保存: {save_path} (原始名: {image_file.filename})") + # + # # 返回字符串路径，generate_img_vector 直接收 str 就行 + # return str(save_path) \ No newline at end of file diff --git a/rag/ragapi.py b/rag/ragapi.py index 355c20f..96b783c 100644 --- a/rag/ragapi.py +++ b/rag/ragapi.py @@ -6,10 +6,13 @@ import traceback import json import math import uuid -from rag.service_opts import get_service_params, sor_get_service_params +import os +from rag.service_opts import get_service_params, sor_get_service_params, sor_get_embedding_mode, get_embedding_mode from rag.rag_operations import RagOperations from langchain_core.documents import Document +REAL_PHOTO_ROOT = "/home/wangmeihua/kyrag/files" + helptext = """kyrag API: 1. 得到kdb表: @@ -134,7 +137,16 @@ async def fusedsearch(request, params_kw, *params): debug(f"params_kw: {params_kw}") # orgid = "04J6VbxLqB_9RPMcgOv_8" # userid = "04J6VbxLqB_9RPMcgOv_8" - query = params_kw.get('query', '') + query = params_kw.get('query', '').strip() + img_path = params_kw.get('image') + if isinstance(img_path, str): + img_path = img_path.strip() + relative_part = img_path.lstrip("/") + real_img_path = os.path.join(REAL_PHOTO_ROOT, relative_part) + if not os.path.exists(real_img_path): + raise FileNotFoundError(f"图片不存在: {real_img_path}") + img_path = real_img_path + debug(f"自动修复图片路径成功: {img_path}") # 统一模式处理 limit 参数,为了对接dify和coze raw_limit = params_kw.get('limit') or ( params_kw.get('retrieval_setting', {}).get('top_k') @@ -189,43 +201,211 @@ async def fusedsearch(request, params_kw, *params): service_params = await get_service_params(orgid) if not service_params: raise ValueError("无法获取服务参数") + # 获取嵌入模式 + embedding_mode = await get_embedding_mode(orgid) + debug(f"检测到 embedding_mode = {embedding_mode}（0=文本, 1=多模态）") - try: - timings = {} - start_time = time.time() - rag_ops = RagOperations() + # 情况1：query 和 image 都为空 → 报错 + if not query and not img_path: + raise ValueError("查询文本和图片不能同时为空") - query_entities = await rag_ops.extract_entities(request, query, service_params, userid, timings) - all_triplets = await rag_ops.match_triplets(request, query, query_entities, orgid, fiids, service_params, - userid, timings) - combined_text = _combine_query_with_triplets(query, all_triplets) - query_vector = await rag_ops.generate_query_vector(request, combined_text, service_params, userid, timings) - search_results = await rag_ops.vector_search(request, query_vector, orgid, fiids, limit + 5, service_params, - userid, timings) + # 情况2：query 和 image 都存在 → 报错（你当前业务不允许同时传） + if query and img_path: + raise ValueError("查询文本和图片只能二选一，不能同时提交") - use_rerank = True - if use_rerank and search_results: - final_results = await rag_ops.rerank_results(request, combined_text, search_results, limit, service_params, + # 3. 只有图片 → 以图搜图 走纯多模态分支 + if img_path and not query: + try: + debug("检测到纯图片查询，执行以图搜图") + rag_ops = RagOperations() + + timings = {} + start_time = time.time() + + # 直接生成图片向量 + img_vector = await rag_ops.generate_image_vector( + request, img_path, service_params, userid, timings, embedding_mode + ) + + # 向量搜索（多取 50 条再截断，和文本分支保持一致） + search_results = await rag_ops.vector_search( + request, img_vector, orgid, fiids, limit + 50, service_params, userid, timings + ) + + timings["total_time"] = time.time() - start_time + + # 可选：搜索完后删除图片，省磁盘（看你需求） + # try: + # os.remove(img_path) + # except: + # pass + + final_results = [] + for item in search_results[:limit]: + final_results.append({ + "text": item["text"], + "distance": item["distance"] + }) + + return { + "results": final_results, + "timings": timings + } + except Exception as e: + error(f"融合搜索失败: {str(e)}, 堆栈: {traceback.format_exc()}") + return { + "records": [], + "timings": {"total_time": time.time() - start_time if 'start_time' in locals() else 0}, + "error": str(e) + } + + if not img_path and query: + try: + timings = {} + start_time = time.time() + rag_ops = RagOperations() + + query_entities = await rag_ops.extract_entities(request, query, service_params, userid, timings) + all_triplets = await rag_ops.match_triplets(request, query, query_entities, orgid, fiids, service_params, + userid, timings) + combined_text = _combine_query_with_triplets(query, all_triplets) + query_vector = await rag_ops.generate_query_vector(request, combined_text, service_params, userid, timings, embedding_mode) + search_results = await rag_ops.vector_search(request, query_vector, orgid, fiids, limit + 50, service_params, userid, timings) - debug(f"final_results: {final_results}") - else: - final_results = [{k: v for k, v in r.items() if k != 'rerank_score'} for r in search_results] - formatted_results = rag_ops.format_search_results(final_results, limit) - timings["total_time"] = time.time() - start_time - info(f"融合搜索完成，返回 {len(formatted_results)} 条结果，总耗时: {timings['total_time']:.3f} 秒") + use_rerank = True + if use_rerank and search_results: + final_results = await rag_ops.rerank_results(request, combined_text, search_results, limit, service_params, + userid, timings) + debug(f"final_results: {final_results}") + else: + final_results = [{k: v for k, v in r.items() if k != 'rerank_score'} for r in search_results] - return { - "records": formatted_results, - "timings": timings - } - except Exception as e: - error(f"融合搜索失败: {str(e)}, 堆栈: {traceback.format_exc()}") - return { - "records": [], - "timings": {"total_time": time.time() - start_time if 'start_time' in locals() else 0}, - "error": str(e) - } + formatted_results = rag_ops.format_search_results(final_results, limit) + timings["total_time"] = time.time() - start_time + debug(f"融合搜索完成，返回 {len(formatted_results)} 条结果，总耗时: {timings['total_time']:.3f} 秒") + + return { + "records": formatted_results, + "timings": timings + } + except Exception as e: + error(f"融合搜索失败: {str(e)}, 堆栈: {traceback.format_exc()}") + return { + "records": [], + "timings": {"total_time": time.time() - start_time if 'start_time' in locals() else 0}, + "error": str(e) + } + +# async def fusedsearch(request, params_kw, *params): +# """ +# 融合搜索，调用服务化端点 +# +# """ +# kw = request._run_ns +# f = kw.get('get_userorgid') +# orgid = await f() +# debug(f"orgid: {orgid}，{f=}") +# f = kw.get('get_user') +# userid = await f() +# debug(f"params_kw: {params_kw}") +# # orgid = "04J6VbxLqB_9RPMcgOv_8" +# # userid = "04J6VbxLqB_9RPMcgOv_8" +# query = params_kw.get('query', '') +# # 统一模式处理 limit 参数,为了对接dify和coze +# raw_limit = params_kw.get('limit') or ( +# params_kw.get('retrieval_setting', {}).get('top_k') +# if isinstance(params_kw.get('retrieval_setting'), dict) +# else None +# ) +# +# # 标准化为整数值 +# if raw_limit is None: +# limit = 5 # 两个来源都不存在时使用默认值 +# elif isinstance(raw_limit, (int, float)): +# limit = int(raw_limit) # 数值类型直接转换 +# elif isinstance(raw_limit, str): +# try: +# # 字符串转换为整数 +# limit = int(raw_limit) +# except (TypeError, ValueError): +# limit = 5 # 转换失败使用默认值 +# else: +# limit = 5 # 其他意外类型使用默认值 +# debug(f"limit: {limit}") +# raw_fiids = params_kw.get('fiids') or params_kw.get('knowledge_id') # +# +# # 标准化为列表格式 +# if raw_fiids is None: +# fiids = [] # 两个参数都不存在 +# elif isinstance(raw_fiids, list): +# fiids = [str(item).strip() for item in raw_fiids] # 已经是列表 +# elif isinstance(raw_fiids, str): +# # fiids = [f.strip() for f in raw_fiids.split(',') if f.strip()] +# try: +# # 尝试解析 JSON 字符串 +# parsed = json.loads(raw_fiids) +# if isinstance(parsed, list): +# fiids = [str(item).strip() for item in parsed] # JSON 数组转为字符串列表 +# else: +# # 处理逗号分隔的字符串或单个 ID 字符串 +# fiids = [f.strip() for f in raw_fiids.split(',') if f.strip()] +# except json.JSONDecodeError: +# # 如果不是合法 JSON，按逗号分隔 +# fiids = [f.strip() for f in raw_fiids.split(',') if f.strip()] +# elif isinstance(raw_fiids, (int, float)): +# fiids = [str(int(raw_fiids))] # 数值类型转为字符串列表 +# else: +# fiids = [] # 其他意外类型 +# +# debug(f"fiids: {fiids}") +# +# # 验证 fiids的orgid与orgid = await f()是否一致 +# await _validate_fiids_orgid(fiids, orgid, kw) +# +# service_params = await get_service_params(orgid) +# if not service_params: +# raise ValueError("无法获取服务参数") +# # 获取嵌入模式 +# embedding_mode = await get_embedding_mode(orgid) +# debug(f"检测到 embedding_mode = {embedding_mode}（0=文本, 1=多模态）") +# +# try: +# timings = {} +# start_time = time.time() +# rag_ops = RagOperations() +# +# query_entities = await rag_ops.extract_entities(request, query, service_params, userid, timings) +# all_triplets = await rag_ops.match_triplets(request, query, query_entities, orgid, fiids, service_params, +# userid, timings) +# combined_text = _combine_query_with_triplets(query, all_triplets) +# query_vector = await rag_ops.generate_query_vector(request, combined_text, service_params, userid, timings, embedding_mode) +# search_results = await rag_ops.vector_search(request, query_vector, orgid, fiids, limit + 50, service_params, +# userid, timings) +# +# use_rerank = False +# if use_rerank and search_results: +# final_results = await rag_ops.rerank_results(request, combined_text, search_results, limit, service_params, +# userid, timings) +# debug(f"final_results: {final_results}") +# else: +# final_results = [{k: v for k, v in r.items() if k != 'rerank_score'} for r in search_results] +# +# formatted_results = rag_ops.format_search_results(final_results, limit) +# timings["total_time"] = time.time() - start_time +# debug(f"融合搜索完成，返回 {len(formatted_results)} 条结果，总耗时: {timings['total_time']:.3f} 秒") +# +# return { +# "records": formatted_results, +# "timings": timings +# } +# except Exception as e: +# error(f"融合搜索失败: {str(e)}, 堆栈: {traceback.format_exc()}") +# return { +# "records": [], +# "timings": {"total_time": time.time() - start_time if 'start_time' in locals() else 0}, +# "error": str(e) +# } # async def text_insert(text: str, fiid: str, orgid: str, db_type: str): async def textinsert(request, params_kw, *params): diff --git a/rag/service_opts.py b/rag/service_opts.py index 9cb7248..559ade5 100644 --- a/rag/service_opts.py +++ b/rag/service_opts.py @@ -94,7 +94,7 @@ async def sor_get_embedding_mode(sor, orgid) -> int: async def get_embedding_mode(orgid): db = DBPools() - debug(f"传入的orgid是：{orgid}") + # debug(f"传入的orgid是：{orgid}") dbname = get_serverenv('get_module_dbname')('rag') async with db.sqlorContext(dbname) as sor: return await sor_get_embedding_mode(sor, orgid) diff --git a/wwwroot/test.ui b/wwwroot/test.ui index 622ccf2..1a5caa5 100644 --- a/wwwroot/test.ui +++ b/wwwroot/test.ui @@ -18,6 +18,11 @@ "editable": true, "rows": 5 }, + { + "uitype": "image", + "name": "image", + "label": "上传查询图片（可选）" + }, { "name": "fiids", "uitype": "checkbox", diff --git a/wwwroot/test_query.dspy b/wwwroot/test_query.dspy index 381290a..6e49ffa 100644 --- a/wwwroot/test_query.dspy +++ b/wwwroot/test_query.dspy @@ -7,14 +7,15 @@ if not orgid: message='请先登录' ) -fiids = params_kw.fiids query = params_kw.query +image = params_kw.image +fiids = params_kw.fiids limit = params_kw.limit -if not query or not fiids or not limit: +if (not query and not image) or not fiids or not limit: return UiError( title='无效输入', - message='请输入查询文本并选择至少一个知识库' + message='请输入查询文本或上传image并选择至少一个知识库和填写返回条数' ) try: