From d2d33c5606d7e08f7c8f9c12191c724556d888e1 Mon Sep 17 00:00:00 2001
From: Mikameel <mikameel@outlook.com>
Date: Mon, 13 Oct 2025 10:18:21 +0800
Subject: [PATCH] num predict low: 16

---
 server/src/ollama_client.rs            |   9 +++++++
 server/src/stream.rs                   |  34 +++++++++++++++++--------
 worker/__pycache__/vlm.cpython-311.pyc | Bin 0 -> 10839 bytes
 worker/vlm.py                          |  21 ++++++++++++++-
 4 files changed, 52 insertions(+), 12 deletions(-)
 create mode 100644 worker/__pycache__/vlm.cpython-311.pyc

diff --git a/server/src/ollama_client.rs b/server/src/ollama_client.rs
index 46ddfe0..52c6ee0 100644
--- a/server/src/ollama_client.rs
+++ b/server/src/ollama_client.rs
@@ -84,6 +84,7 @@ pub async fn send_to_ollama(
     prompt: String,
     model: String,
     ollama_host: String,
+    num_predict: Option<i32>,
 ) -> Result<mpsc::Receiver<Result<Bytes, reqwest::Error>>, reqwest::Error> {
     tracing::info!("Sending images to Ollama: {:?}", images_batch.len());
     let images_b64: Vec<String> = images_batch
@@ -96,6 +97,14 @@ pub async fn send_to_ollama(
         "images": images_b64,
         "model": model.clone(),
         "stream": true,
+<<<<<<< Updated upstream
+<<<<<<< Updated upstream
+=======
+        "num_predict": num_predict.unwrap_or(16),
+>>>>>>> Stashed changes
+=======
+        "num_predict": num_predict.unwrap_or(16),
+>>>>>>> Stashed changes
     });
 
     let url = format!("{}/api/generate", ollama_host);
diff --git a/server/src/stream.rs b/server/src/stream.rs
index b7d8bbe..463cae8 100644
--- a/server/src/stream.rs
+++ b/server/src/stream.rs
@@ -2,6 +2,7 @@ use actix_web::{rt, web, Error, HttpRequest, HttpResponse};
 use actix_ws::Message;
 use futures::{select, FutureExt};
 use futures_util::StreamExt as _;
+use serde_json::Value;
 use tokio::time::{self, Duration, Instant};
 
 use crate::{config, ollama_client::{send_to_ollama}};
@@ -12,10 +13,11 @@ async fn proxy_ollama_response(
     prompt: String,
     model: String,
     ollama_host: String,
+    num_predict: i32,
 ) {
     let mut session_clone = session.clone();
     rt::spawn(async move {
-        match send_to_ollama(images_batch, prompt, model, ollama_host).await {
+        match send_to_ollama(images_batch, prompt, model, ollama_host, Some(num_predict)).await {
             Ok(mut rx) => {
                 while let Some(res) = rx.next().await {
                     match res {
@@ -43,7 +45,7 @@ async fn handle_binary(
     images: &mut Vec<Vec<u8>>,
     bin: bytes::Bytes,
     session: &mut actix_ws::Session,
-    last_prompt: &Option<String>,
+    last_prompt: &Option<(String, i32)>,
     model: String,
     ollama_host: String,
     image_batch_size: usize,
@@ -51,9 +53,9 @@ async fn handle_binary(
     images.push(bin.to_vec());
 
     if images.len() >= image_batch_size {
-        if let Some(prompt) = last_prompt.clone() {
+        if let Some((prompt, num_predict)) = last_prompt.clone() {
             let images_batch = std::mem::take(images);
-            proxy_ollama_response(session, images_batch, prompt, model, ollama_host).await;
+            proxy_ollama_response(session, images_batch, prompt, model, ollama_host, num_predict).await;
         } else {
             let _ = session.text("No prompt received for image batch".to_string()).await;
         }
@@ -64,16 +66,24 @@ async fn handle_text(
     session: &mut actix_ws::Session,
     images: &mut Vec<Vec<u8>>,
     text: String,
-    last_prompt: &mut Option<String>,
+    last_prompt: &mut Option<(String, i32)>,
     model: String,
     ollama_host: String,
 ) {
-    *last_prompt = Some(text.clone());
+    let (prompt, num_predict) = if let Ok(value) = serde_json::from_str::<Value>(&text) {
+        let prompt = value["prompt"].as_str().unwrap_or(&text).to_string();
+        let num_predict = value["num_predict"].as_i64().map(|n| n as i32).unwrap_or(16);
+        (prompt, num_predict)
+    } else {
+        (text.clone(), 16)
+    };
+
+    *last_prompt = Some((prompt.clone(), num_predict));
 
     // If prompt updated, send the images to Ollama
     if !images.is_empty() {
         let images_batch = std::mem::take(images);
-        proxy_ollama_response(session, images_batch, text, model, ollama_host).await;
+        proxy_ollama_response(session, images_batch, prompt, model, ollama_host, num_predict).await;
     }
 }
 
@@ -92,7 +102,7 @@ pub async fn ws_index(req: HttpRequest, stream: web::Payload, vlm_config: web::D
 
     rt::spawn(async move {
         let mut images: Vec<Vec<u8>> = Vec::new();
-        let mut last_prompt: Option<String> = None;
+        let mut last_prompt: Option<(String, i32)> = None;
 
         let mut inference_interval = time::interval_at(Instant::now() + Duration::from_secs(30), Duration::from_secs(10));
         inference_interval.set_missed_tick_behavior(time::MissedTickBehavior::Skip);
@@ -141,9 +151,11 @@ pub async fn ws_index(req: HttpRequest, stream: web::Payload, vlm_config: web::D
                         tracing::error!("Error sending ping: {:?}", e);
                         break;
                     }
-                    if let Some(prompt) = last_prompt.clone() {
-                        tracing::info!("Inference interval fired: running handle_text with last prompt");
-                        handle_text(&mut session, &mut images, prompt.clone(), &mut last_prompt, model.clone(), ollama_host.clone()).await;
+                    if let Some((prompt, num_predict)) = last_prompt.clone() {
+                        if !images.is_empty() {
+                            let images_batch = std::mem::take(&mut images);
+                            proxy_ollama_response(&mut session, images_batch, prompt, model.clone(), ollama_host.clone(), num_predict).await;
+                        }
                     }
                 }
             }
diff --git a/worker/__pycache__/vlm.cpython-311.pyc b/worker/__pycache__/vlm.cpython-311.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..b7d36f514592c24de4d6383a57a90bc875da3f52
GIT binary patch
literal 10839
zcmdT~Yitu)mM+^>cDd}tj_te%Bt>`{NJ7Yq4kVB_frMs4pveO~CRIrs96McANx;~x
zOshwmK}elONMu=N+Ot|QL9kLZ>eldMS7Nv6Zf3Q!Q-un(RoX~ME3|6uu9V*HmRha-
zvFBFB<wrtKW?#~lt4`hby>;$A=brOj|0*Yk!r=P1K4pBR1jGIbMUt0!E%5bs5V(Ue
z7-I-xgWB6LXi($EL8JOM4Vu(9K8VBH7$hdmgXT%gpk<OABn_yJDQKOf1}Qa+2W^w~
zLAwDnVl2fF_YD~K5FYIs%z;!pYnaGQro<h|Q2LuOGnV^Cs0ojDJtDJOvN6{8utBFf
zGBi>OVtI@WsEhG1xvZOUu%1zzalS_k<}-PaUck737BX(2MU3Gr=FR^SHGbM_Qu0Ov
zp#VSbo8rPFEYAl*V+uLy4+MP^;meABBs@74WCa$*yat6DV+CI@JT}I15Hq(018hi8
zOf$hr#r9@TPfKr$@5GyD`y&P8f-u!sRTT`6_=DqNUTCbYuC1wsIyj1r=xl*Oknot`
z`5MtXm=zNslDsDLl5w3O$vpIC42&_8T`@yOeblmg#qne){T($hrmXr#LGKg8CWvI|
zsFA@R5t(HTm|#!lV_4jh3a3K%A^1>_#!P}!k6>|YDx3;2W+6Wni<&a0KLS~qDX{n~
z?_p6BV=C6GwPRWR;!!M$&tP2HCi*9%*oM7W&tjXXO>JB|gHtwrplN~Vy6@UVZ?=tl
zV_>X1Frg^b{7#JPjS(3w=A?2{A;zw?$mC?qLY6oO43Bcxt%Yr=Rwi`eW*phGHGnbJ
z81W8)#;Ap64YSx)(*QPW@H!(i6b)BLh@a+Ip7zs|VTKLT0SL~3So$vq*#q>{bT9~#
zN(TfwFiMA5B)NX>8XX7;VY<Dgzs1+y-B+d7<SVuIt_B32o(%}&^awOaSEGtZ&RIq)
zppeOW3eNZizhc&^_k+-q?9~7-_`+8ru9oB=v~lr`bcQe^A#=m?bt#ZL*ccYYE>bZA
zh#8hzoF5o(+CT>x6hiId0p>L-#>#_S9@OEHnA7z4IkREz3d>c2YN?#M_C0kp2qQq+
zs5!EaRe9r6EBkb=R325e4c`obWLY!{ek+hUOf2cv?xpF)j`*H?r}c!dAs>qr_o{sM
z2RYWyTuU*#kv<*SqUD9c0zDd@4l()!G}4h0Ef38I%h3Gv2%3vgh+Tusf=-pgG|UPr
zpz*AcRtT2k!rTMk0DD#7{9huC<8>;=Ft1qH&`f{}hZGa&S_Pl-3*!nt5e|eDa?*c=
zWda<pn8U%Kf6}ku=tC8=rjQjp5E>0DCSKs68uX@EJFbqfQ$ipd;tHVy$hyWWILxKs
zH7k_HzL0;CRY)}9K4^L^l}BL??No|W=dZ?ZUNNvd+C$MIrD@KC3KeWTBU}|^B0=CU
z!UxV_t5gBudF9T(-FT<@R&#vc(x~LFlihXm<Vt?w?YHg>-x`kZT{<D<*U9;H^VAAS
z-Lx&7T=Ixyr9@WBWF-W?B{5ggoswH6A8fn5ZQi^>=Du&cVG})j|FULj=E0FiM<h?Z
zL^jA|gGe?c)H#1_etcCvd;;cMD(RLXi7a|)%NL8iq7B_EcGvr-Zk!T}YL+LYqT`ai
zMYgwy_LhX*HQ)Iv+5W-hM#<A8k<BvMERxM@{Vtr8$TFEM6Uj1tB-P6&B~P<N9+Syq
zV)Fj(mCdt?8HqH4D%fw>#r%q8O3H71((t=B@yw5Y+xhpMV(Zy?yF~WOWWPxEuaLHz
z<U-!hY&UK58Nct~S|y-Oc=I%{`Q0{q`)=$LXNw13Pj^!tO{S*@TZ|otO`jenAhJ$U
z)ou4K)C4;SHqs!5B=rw%x!GzuW)#vkz6pJsWlYiRt=SYcarseGT;EioYyD#IOhpOG
zkpL}Wj{NJI0s}!A0UhS2&z<XTpQ)t-3_Z$)Cux6(4ov#TSQ^O+J&jBlnq$Y<t5ce0
z>j{Ui@F2Kcsw$KeSPtriXknbCM?noDK}f1fKQek6>GCQh7gd4K)U;4FH69kiK8<qh
z+rH|Cy4u6l^#>d3YisI!_0<iR-@bOpcj(}ux`T(S5BjR>>uPHcCete?rpC(Yie|ce
ztzbF57xZsH*pvDmulkuB(toET=l2fzD@I!?IuBO7aq;@$SVelYHdaxa9<PDy!*BR6
zx3pf&PEKp2`cTZvZGqMm?17mBvl!!$dQ`~EeqQ~0#mKRWnP>gn$hbm`apCEy2e?A1
zy!9d8s+GVxT;mp~Z@(}y&Z8Mm>4MxTKgY8^tyX~HjzDfDJp2V9pabj}kvrFoehEk3
z&Dk#;WzQXDaY}ORmmT}(PQ0?`-t110{sig#iZJDvzrldab*z#YnX`%+h|+|^^ZxZ4
z*B7gQ_Rh_BL~>igmNS2dD}cuTe=cu72D&kGg`*`Y1SVNtfQ7iZHQvt!z*vB72CR7o
z)_HP)2Lo4Fy4-hNZ!YF5r$G`jk}Avp*99RZ09g`Jp1$f!eS43cIV$kY!wjr>AJF=k
zcki)AUv#)Ky3?x|1)i&hndS~5(f}lD@#Uv4KK+}yW+-_Kl|Bn(lWR~cXn0&C*;{3M
z>s&{|Vn>T_dvfvBN)B3f&>~5vmft_${Zn-P_Sl!m`BZEIJ2}O2egVul#TJ^L1P=>#
z?-5}nUHc&iK-Izrfg*1-M!;jvaA9<JWYaE?MM!$N(0p9i$Pm#l(9RZQ<TE&;LYhIo
zk;B0d>q=#5Ym}d<zri_5Gq9tUb-90djOO%l3A#y*VX(6RS03t|$&9JrrG8y{{p>m$
zjhG>0WYpM>UHm9!-2@M<QR_ynWH=ccx_D6wxMS8aLxziy<>*CmhPr2bWV;XZ_fVh1
z7!@TMyMXMuwJU1f;Olc#9@r5TQieD5`NL(S)Pz1~Y74oDMj$OlZh{|<41QDV`^ro)
zU^hSBM61s1T1;M)3?LN4by~5W@9FjRzS-W<6CuGUp7A#}T#n?OQ;h(~N?P0zm*)GV
z31w3x4@?J!P7}?hNLk->D5QE;p;4A&LnB}YXj|C`82&<2Bxfy(_CvlY;tntebUkx`
zodGz6inz|`Sc+D<k$zz0T7Z*WD-bWPdOC_3AeYH0u;0}V!152PwmSd4p*&(NKL8Ij
z)twv%9iA4z4gsgaok~l>GYq%_;W6HutL^K^<Kd7u%I!d8CnCELK?b>E0%M2UjzZ)Z
z8$v3IRfy4G*e|G?wW=)@5}ePeaENEUHiZHw4B<mCSgsS*v!!eTKJ{IS4K^%J(6)ZX
z%7&O^5Y<=R4aEY@P6t7aA67>KF&?QNWIn82nn#htO^1Bo&uDYNorf~L@bD6d{2Z2W
z70jJVI6ZUS-<q*J_igG<?ycO#>q{pkXM^l)nCpIpQ-5;m-BY4-M?4@otHeW{lC$%7
z?q@?%_ki3zAUOxb;jrWkzu=}<F>t$EjbB0IhBo6j+6!WNZ77XsM@c}T7*ObrC|P_%
z!gtB|E)n0Q_f!3G+h3a=pOOxr08c}z?Url1C1>}uHznuLi%Wj-@}zVrBwq?i&X9zM
zWjrk6;Y9AvxvrN5+gC9=QKu5*@+94vC2%+DB(h#6>qWADCD-$Q_(nM4DoMDvB|N1I
zM_-lgj1Ng=)pA*NqHNzPVam^4#SpP#j?%9%(m{QL$h>LQg;|{oH9s$1qNL&)xwuAh
z*DN!VyYcZ^$$k7q+XZo8L~3K?Hb`T{u`81MiWquZa=$GKv+(KDh8DyJPYdEeyRlo1
zsMOkRoVTO~4E`7Y;P9ivlBY%@Yh|)lRBxR_yCi(KjPDll-K%EI>R7O>nv8Zc@`=ov
z^`@&hm}j&#sN3Q%V7~7Fi0~706kL=m$T-Hht?Aoi97?Re02PGv)CPe7+Ot^H5H%v4
z?4Bux1J0PX5l^7r)C=KcD$)lj5B1a-5hc=G8}k)4hx<{+NYdI%wgwxlC1V>*YbA4p
zs=6f+g`RYW3dJ+a!M2*0*19#l^q*lr&-(6nGtki2;Eh>|TH_U<XK6?!pl6ntIck0v
zfSj259djCDWjn3fWD(C>?B#KAwdiwwJqWG>^GCP_0J9mOV;?vIXk$@v2<|w90P?{n
zcz`)9NPU8C`|=s2Ms}tE6L3lxdRA+*kp|TwO!Gd3T{T50@b5J3)D$UtU5!X_azN0M
zkb%s2DOE8lmRUcCv^~;xZ%>204e-bm;M1xmOLnJ?8L3YN7Z6;t2ZE|j^;#6$+5mmi
zT#!Q^PU<ri%as2bXjO(gjnY*g2+eDTp0lH`^sP0*jX?GkJp8`{0X1iFiIy#kSC;yo
zmsW|TRY=%!Z{&)F6*?S}Aam!<AW-gNP=|@Uf>jId0HFdh*R@KKgmcARa=R&BvfS{b
zRdjbr?k?HgHBTmNZrN5AuX%3UE!uV?1f=Iik63){(;ZL7<<3E=bx3X<l8T4KOA}J@
zgg7}ZPe!ETh-AMm+pmlE>npAjack%D)sJHEf?INR$*wNZ-t|fcQzUz>Y_AoQ_jfA@
z05K5G6iMVd<}E6SVS;o7Vj#HoQEN`?5$tb|P;GY8Cxo%hV*12FKsdv@Lz0Fh@az9j
z^}`qd7mNXbh%{Z0f-$lI3o!5gAqG&y#^&^WXf_S6$6o_(7^8S5{);pO_sb~p8gL`a
zxC2=-F(B#jG$R-~Ub`~K@Ka;Tf{R(wHBL%T#L+3rx(jq!)?pHuacQMCQ9t{1#0sXX
zX-)g(XYgA$bfeZNnaLXiXx_MbB$w8?8ibNrdiv(?p=_EvlB$u(71Sn}fH<6CMR4R^
zMhKwn7(}K*V-e?Y<z7gPqNFG!cvWD<Yu7Lg3j$6N-bn<*vw+3cLN8n$BK1HN69}aW
zj&NltY6e3P!K7yY)D+|>I3MN&FQr&jWJtwI&!BSDj6WD)e4rzE?hQ3TZQX||bB$WV
zR|Pm!3$p$Y%c$w87_umpoG=#{3xxbZ6>#AZ&(`z>yfo9uhFpBjvf?=o3cLpo|1Ut0
zVdMeBC~x1~$yc6&+b0*VNgl84@vdS{qCzDg`uW@LT)K5h%HJjD@0zDH1Ij9HsaT3g
z&PLhUI8P*OAf+~IMZ#5hXaBAJi~bKPZ&%K@y)4-g?}?xL)xe(*{Br0ohT=VPN%efs
zO3Ai+4Idu8dvwXVd{inqE|(mi?}1jgmEG(5u;*^iQt|Q$X={_bwQ0UL;V4<W`rNTc
zbnIDix^7z+E0(rO&U)EdFFNaA+6%w1mp-?bE=DE$0oi^)v>*8HCGPkF_dLfv5?&zV
z1tMPX(uU}IlR@{S&wz3OlK~V$n%Lm!OEudQgkvFnqet6s=8|NLYi7q0A<aaIr_}(~
zuU!c9+-S>YaE4&a_l)4#>tX^up;1~=ql}vGL#rvHL)(bKP)QSqD88}EU`PSs^=sFJ
zt`1&P+R9LwBAOM;5K2%w=uBtrdhG~of7A%<OiosxVBDony))+k)K>bDD9+@jaSdl2
zX)yv66#2+`A6CIbv`wg2%o_SfFlZ1;xORa-^ZOW;F{5cmxye#szU5`kk$`~HwJVxs
zI595p*WH<41#{^6!8-qiRRJp^?WhsV)%>*Y&sb5g#h5ek0|OT7UdM@}o8m;lf0q+*
z@QKB!cc>7ly;DYHG6xyy*?9ue!AXv1ON=#JVm8m4sFlIi@rGZ=o6-Lp-Y|v#)2rkM
zdAdEx(e+wr-TH+&@fMv{2z9qpx6w$+>9B@upmP;C5CxC~9g?7}uqm(r$|K^)DuL{M
zWV|aDz_ocetx=5N6STlE6bFKE`a62hbo90KpX=-35LBxoC0f?$bG^PZeI4!HZT+h2
zz`cp`I}zys67h7P!&Ex;2NoLrS;eTC;|M6@kX`vDgF=2(A?;K)-O1pFAk&+t?y1}<
zRC)&@!-ycw&mjatgH_bi6NIuTCOC3e;R($xP`lfO%8ej`HVF=CAwW^!^Fwf|3uEK1
zK#D?XfR1|XYF63DbE+9(0)r%ls&M_N4b^)<+d#@1SThGqfRu1Qft+8!!+!%x6wH7e
z=R)Dl;knKg*S5ug<k~B{_RgJpWufMeNR~p`Qn*<C+_F`)Y}GxA+U09f;c<y<k;xX3
zY<cM^Ocd?~RKY<3s(@ewICr(?EqE6A#oF7~;yWaAw@mI9$=#X*f;$MuN<qnD=Z7cn
zo{YynzWmX|Z$iHgi5=&~3qw-<uv|YZZM`UOy(kr2k_#@)+g1vS7k7WS@9w_%#gAJ*
zI{BN^zdkLt^^4~(NVNlU?SQmpP~I{q6%5G*Lr~aN7B|PEGF`ho`uKue-}%fY`h%+&
zcHPj4U<r2@LM4EHqtbZ`xD~3qVIv%1$gC0ct?GH}{>LLvcK%i1Mc+B`{997r1-b8n
zSTg`OI&>YB$w84E1O}`WZdt7UVCMGBX4tT5G+F(I6?@TQK|Dt;u3R=Pw|``PT>Hc=
z**j!=hp67)tvHIn|ACU}RCm9spR_&0pLV~vFeDCNlrCJ7FI*De@WDObBiVl}+kY&s
zyVu+&Xc0m$ku#Y5sX2x4@eufS3q}+yFpF(*{9bo_R>G@fyh_BYR&guL8|hg%zt}71
z?^|kH#vgSr&pyE=vO^|2MD^BKX+c~}5}lP5PfSnRpITq^za?H6kopJZ{z0*B2yVp6
zVVN8j$>BeimFkMh_zDN~i@w5z5&9C8@a;0bT}<BmMOgiRv!}JZox+|{RM!sEQ%_4t
zSE=dKgQcAg<7c@B2!H0Fy7Gz7iVt-;iO+2Wq<roqAm#IX6AG6SQ0?>P(i7XTXWOWg
zIi_cOT1rlmroSTzh=4*-_0n<Z$Xm<}z)J&LRo`JRilQ-~|7u+3HNjVv29kayNBfHg
zz!#y>z!-;|J{5uG5Fntb8Kg}#<$^%k6cVdxo$DwKp`e;xL1w=y^qdQYkX@rX2j7!N
z0cV{Kvd!FkkOqnv<BtPfH5m+s1ZJC4e+kSwr~c9c)SUWDV2(NUm%tju*WU?jr<i#s
zuwpUu22_Ib%x_ufkgQu|>y|n5ORHnPV4?5!Ao4%ry<*KNDZfX~?-8xNlC@X1_Rg7C
z@e%{T1UhNN43sw1Ror0!gi|Nwb`;gg*6jv>J%BtR)R#DTQGI8VPJ_z;&W%p?78wdx
RF`d-zH<Yhp>xs(Y{{)4rbMOEF

literal 0
HcmV?d00001

diff --git a/worker/vlm.py b/worker/vlm.py
index 33a666a..e2d022a 100644
--- a/worker/vlm.py
+++ b/worker/vlm.py
@@ -74,18 +74,25 @@ def parse_image_timestamp(image_path):
         return match.group('ts')
     return ""
 
-def run_inference(vlm_prompt, prompt, image_paths):
+def run_inference(vlm_prompt, prompt, image_paths, temperature: float = 0.2, num_predict: int = 16):
     start_image = None
     end_image = None
     
     # Get model names from environment variables
     vlm_model = os.environ.get("VLM_MODEL", "llava:7b")
+<<<<<<< Updated upstream
     llm_model = os.environ.get("LLM_MODEL", "llama3:latest")
     
     logger.info("Using VLM model: " + vlm_model)
     logger.info("Using LLM model: " + llm_model)
     
     # Ensure models are available
+=======
+
+    logger.info("Using VLM model: " + vlm_model + " with temperature=" + str(temperature) + " and num_predict=" + str(num_predict))
+
+    # Ensure VLM model is available
+>>>>>>> Stashed changes
     ensure_model_available(vlm_model)
     ensure_model_available(llm_model)
 
@@ -98,6 +105,13 @@ def run_inference(vlm_prompt, prompt, image_paths):
             model=vlm_model,
             prompt=vlm_prompt,
             images=[image_path],
+<<<<<<< Updated upstream
+=======
+            options={
+                "temperature": float(temperature),
+                "num_predict": num_predict
+            },
+>>>>>>> Stashed changes
         )
         results += '"' + parse_image_id(image_path) + '",' + '"' + parse_image_timestamp(image_path) + '",' + '"' + res.response + '"\n'
         logger.info("Inference output: " + str(res))
@@ -183,7 +197,12 @@ def run(conn, job, input_dir, output_dir):
         return
 
     try:
+<<<<<<< Updated upstream
         results = run_inference(inputs['vlm_prompt'], inputs['prompt'], image_paths)
+=======
+        temperature = float(inputs.get('temperature', os.environ.get("VLM_TEMPERATURE", 0.2)))
+        num_predict = int(inputs.get('num_predict', os.environ.get("VLM_NUM_PREDICT", 16)))
+        results = run_inference(inputs['vlm_prompt'], inputs['prompt'], image_paths, temperature=temperature, num_predict=num_predict)
     except Exception as e:
         logger.error("Error processing job", extra={"job_id": job['id'], "error": str(e)})
         err = {