Partial implementation of dual-mode templates

This commit is contained in:
Cyber MacGeddon 2025-01-10 16:18:30 +00:00
parent 08d24bf5aa
commit 91f2c77abf
21 changed files with 633 additions and 304 deletions

View file

@ -5,6 +5,11 @@ local prompts = import "prompts/mixtral.jsonnet";
{
with:: function(key, value)
self + {
["claude-" + key]:: value,
},
"claude-max-output-tokens":: 4096,
"claude-temperature":: 0.0,
@ -31,48 +36,18 @@ local prompts = import "prompts/mixtral.jsonnet";
.with_limits("0.5", "128M")
.with_reservations("0.1", "128M");
local containerRag =
engine.container("text-completion-rag")
.with_image(images.trustgraph)
.with_command([
"text-completion-claude",
"-p",
url.pulsar,
"-x",
std.toString($["claude-max-output-tokens"]),
"-t",
"%0.3f" % $["claude-temperature"],
"-i",
"non-persistent://tg/request/text-completion-rag",
"-o",
"non-persistent://tg/response/text-completion-rag",
])
.with_env_var_secrets(envSecrets)
.with_limits("0.5", "128M")
.with_reservations("0.1", "128M");
local containerSet = engine.containers(
"text-completion", [ container ]
);
local containerSetRag = engine.containers(
"text-completion-rag", [ containerRag ]
);
local service =
engine.internalService(containerSet)
.with_port(8000, 8000, "metrics");
local serviceRag =
engine.internalService(containerSetRag)
.with_port(8000, 8000, "metrics");
engine.resources([
envSecrets,
containerSet,
containerSetRag,
service,
serviceRag,
])
},