Touchups for pr tool calling

* branch naming convention * hallucination minimization
pluralsh · Dec 4, 2024 · 4acbfba · 4acbfba
1 parent d05fc9f
commit 4acbfba
Show file tree

Hide file tree

Showing 25 changed files with 399 additions and 95 deletions.
diff --git a/assets/src/components/ai/chatbot/AISuggestFix.tsx b/assets/src/components/ai/chatbot/AISuggestFix.tsx
@@ -1,4 +1,10 @@
-import { Button, Markdown, PrOpenIcon, Toast } from '@pluralsh/design-system'
+import {
+  Button,
+  LinkoutIcon,
+  Markdown,
+  PrOpenIcon,
+  Toast,
+} from '@pluralsh/design-system'
 import {
   Dispatch,
   ReactNode,
@@ -8,7 +14,6 @@ import {
   useRef,
   useState,
 } from 'react'
-import { useTheme } from 'styled-components'
 import {
   AiDelta,
   AiInsightFragment,
@@ -87,37 +92,15 @@ function FixPr({
   const [mutation, { data, loading, error }] = useAiFixPrMutation({
     variables: { insightId, messages: [{ role: AiRole.User, content: fix }] },
   })
-  const theme = useTheme()
 
   // TEMP FIX, implement permanent solution in DS
-  const successToastRef = useRef<HTMLDivElement>(null)
   const errorToastRef = useRef<HTMLDivElement>(null)
   useEffect(() => {
-    if (successToastRef.current) successToastRef.current.style.zIndex = '10000'
     if (errorToastRef.current) errorToastRef.current.style.zIndex = '10000'
   }, [data?.aiFixPr])
 
   return (
     <>
-      {data?.aiFixPr && (
-        <Toast
-          ref={successToastRef}
-          severity="info"
-          position="top-right"
-          margin="large"
-          heading="PR Created!"
-        >
-          <a
-            href={data?.aiFixPr?.url}
-            target="_blank"
-            rel="noreferrer"
-            style={{ textDecoration: 'none', cursor: 'pointer' }}
-            color={theme.colors['action-link-inline']}
-          >
-            {data?.aiFixPr?.url}
-          </a>
-        </Toast>
-      )}
       {error && (
         <Toast
           ref={errorToastRef}
@@ -129,13 +112,27 @@ function FixPr({
           {error.message}
         </Toast>
       )}
-      <Button
-        startIcon={<PrOpenIcon />}
-        onClick={mutation}
-        loading={loading}
-      >
-        Open PR
-      </Button>
+      {data?.aiFixPr ? (
+        <Button
+          primary
+          type="button"
+          endIcon={<LinkoutIcon />}
+          as="a"
+          href={data?.aiFixPr?.url}
+          target="_blank"
+          rel="noopener noreferrer"
+        >
+          View PR
+        </Button>
+      ) : (
+        <Button
+          startIcon={<PrOpenIcon />}
+          onClick={mutation}
+          loading={loading}
+        >
+          Create PR
+        </Button>
+      )}
     </>
   )
 }

diff --git a/assets/src/generated/graphql.ts b/assets/src/generated/graphql.ts
@@ -330,11 +330,15 @@ export type AnthropicSettings = {
   __typename?: 'AnthropicSettings';
   /** the anthropic model version to use */
   model?: Maybe<Scalars['String']['output']>;
+  /** the model to use for tool calls, which are less frequent and require more complex reasoning */
+  toolModel?: Maybe<Scalars['String']['output']>;
 };
 
 export type AnthropicSettingsAttributes = {
   accessToken?: InputMaybe<Scalars['String']['input']>;
   model?: InputMaybe<Scalars['String']['input']>;
+  /** the model to use for tool calls, which are less frequent and require more complex reasoning */
+  toolModel?: InputMaybe<Scalars['String']['input']>;
 };
 
 /** a representation of a kubernetes api deprecation */
@@ -635,6 +639,8 @@ export type AzureOpenaiAttributes = {
   endpoint: Scalars['String']['input'];
   /** the exact model you wish to use */
   model?: InputMaybe<Scalars['String']['input']>;
+  /** the model to use for tool calls, which are less frequent and require more complex reasoning */
+  toolModel?: InputMaybe<Scalars['String']['input']>;
 };
 
 /** Settings for configuring against Azure OpenAI */
@@ -644,6 +650,9 @@ export type AzureOpenaiSettings = {
   apiVersion?: Maybe<Scalars['String']['output']>;
   /** the endpoint of your azure openai version, should look like: https://{endpoint}/openai/deployments/{deployment-id} */
   endpoint: Scalars['String']['output'];
+  model?: Maybe<Scalars['String']['output']>;
+  /** the model to use for tool calls, which are less frequent and require more complex reasoning */
+  toolModel?: Maybe<Scalars['String']['output']>;
 };
 
 export type AzureSettingsAttributes = {
@@ -689,6 +698,8 @@ export type BedrockAiAttributes = {
   modelId: Scalars['String']['input'];
   /** aws secret access key to use, you can also use IRSA for self-hosted consoles */
   secretAccessKey?: InputMaybe<Scalars['String']['input']>;
+  /** the model to use for tool calls, which are less frequent and require more complex reasoning */
+  toolModelId?: InputMaybe<Scalars['String']['input']>;
 };
 
 /** Settings for usage of AWS Bedrock for LLMs */
@@ -698,6 +709,8 @@ export type BedrockAiSettings = {
   accessKeyId?: Maybe<Scalars['String']['output']>;
   /** the bedrock model to use */
   modelId: Scalars['String']['output'];
+  /** the model to use for tool calls, which are less frequent and require more complex reasoning */
+  toolModelId?: Maybe<Scalars['String']['output']>;
 };
 
 export type BindingAttributes = {
@@ -3879,13 +3892,17 @@ export type OllamaAttributes = {
   /** An http authorization header to use on calls to the Ollama api */
   authorization?: InputMaybe<Scalars['String']['input']>;
   model: Scalars['String']['input'];
+  /** the model to use for tool calls, which are less frequent and require more complex reasoning */
+  toolModel?: InputMaybe<Scalars['String']['input']>;
   url: Scalars['String']['input'];
 };
 
 /** Settings for a self-hosted ollama-based LLM deployment */
 export type OllamaSettings = {
   __typename?: 'OllamaSettings';
   model: Scalars['String']['output'];
+  /** the model to use for tool calls, which are less frequent and require more complex reasoning */
+  toolModel?: Maybe<Scalars['String']['output']>;
   /** the url your ollama deployment is hosted on */
   url: Scalars['String']['output'];
 };
@@ -3897,12 +3914,16 @@ export type OpenaiSettings = {
   baseUrl?: Maybe<Scalars['String']['output']>;
   /** the openai model version to use */
   model?: Maybe<Scalars['String']['output']>;
+  /** the model to use for tool calls, which are less frequent and require more complex reasoning */
+  toolModel?: Maybe<Scalars['String']['output']>;
 };
 
 export type OpenaiSettingsAttributes = {
   accessToken?: InputMaybe<Scalars['String']['input']>;
   baseUrl?: InputMaybe<Scalars['String']['input']>;
   model?: InputMaybe<Scalars['String']['input']>;
+  /** the model to use for tool calls, which are less frequent and require more complex reasoning */
+  toolModel?: InputMaybe<Scalars['String']['input']>;
 };
 
 export enum Operation {
@@ -9364,6 +9385,8 @@ export type VertexAiAttributes = {
   project: Scalars['String']['input'];
   /** optional service account json to auth to the GCP vertex apis */
   serviceAccountJson?: InputMaybe<Scalars['String']['input']>;
+  /** the model to use for tool calls, which are less frequent and require more complex reasoning */
+  toolModel?: InputMaybe<Scalars['String']['input']>;
 };
 
 /** Settings for usage of GCP VertexAI for LLMs */
@@ -9375,6 +9398,8 @@ export type VertexAiSettings = {
   model?: Maybe<Scalars['String']['output']>;
   /** the gcp project id to use */
   project: Scalars['String']['output'];
+  /** the model to use for tool calls, which are less frequent and require more complex reasoning */
+  toolModel?: Maybe<Scalars['String']['output']>;
 };
 
 export type VerticalPodAutoscaler = {

diff --git a/assets/src/helpers/client.ts b/assets/src/helpers/client.ts
@@ -25,6 +25,7 @@ export const authlessClient = new ApolloClient({
 })
 
 function maybeReconnect(socket, absintheSocket) {
+  const chan = absintheSocket.channel
   console.log('socket reconnect attempt', socket)
   if (socket.connectionState() === 'closed') {
     console.log('found dead websocket, attempting a reconnect')
@@ -37,12 +38,18 @@ function maybeReconnect(socket, absintheSocket) {
       socket.reconnectTimer.scheduleTimeout()
     }
 
-    absintheSocket.channel.rejoin()
-    absintheSocket.channel.rejoinTimer.scheduleTimeout()
     return
   }
 
-  console.log('socket not dead, ignoring')
+  const state = chan.state
+  if (state === 'closed' || state === 'errored') {
+    console.log('broken absinthe channel, rejoining')
+    chan.rejoin()
+    return
+  }
+
+  console.log('found healthy channel', chan)
+  console.log('absinthe socket not dead, ignoring')
 }
 
 export function buildClient(gqlUrl, wsUrl, fetchToken) {
@@ -77,6 +84,10 @@ export function buildClient(gqlUrl, wsUrl, fetchToken) {
   const socketLink = createAbsintheSocketLink(absintheSocket)
   const gqlLink = errorLink.concat(httpLink)
 
+  absintheSocket.channel.onClose(() => {
+    absintheSocket.channel.rejoin()
+  })
+
   const splitLink = split(
     (operation) => hasSubscription(operation.query),
     socketLink,
@@ -105,7 +116,7 @@ export function buildClient(gqlUrl, wsUrl, fetchToken) {
   })
 
   socket.onClose(() => maybeReconnect(socket, absintheSocket))
-  setInterval(() => maybeReconnect(socket, absintheSocket), 10000)
+  setInterval(() => maybeReconnect(socket, absintheSocket), 5000)
 
   return { client, socket }
 }

diff --git a/charts/controller/crds/deployments.plural.sh_deploymentsettings.yaml b/charts/controller/crds/deployments.plural.sh_deploymentsettings.yaml
@@ -81,6 +81,10 @@ spec:
                         - key
                         type: object
                         x-kubernetes-map-type: atomic
+                      toolModel:
+                        description: Model to use for tool calling, which is less
+                          frequent and often requires more advanced reasoning
+                        type: string
                     required:
                     - tokenSecretRef
                     type: object
@@ -122,6 +126,10 @@ spec:
                         - key
                         type: object
                         x-kubernetes-map-type: atomic
+                      toolModel:
+                        description: Model to use for tool calling, which is less
+                          frequent and often requires more advanced reasoning
+                        type: string
                     required:
                     - endpoint
                     - tokenSecretRef
@@ -158,6 +166,10 @@ spec:
                         - key
                         type: object
                         x-kubernetes-map-type: atomic
+                      toolModelId:
+                        description: Model to use for tool calling, which is less
+                          frequent and often requires more advanced reasoning
+                        type: string
                     required:
                     - modelId
                     type: object
@@ -196,6 +208,10 @@ spec:
                         - key
                         type: object
                         x-kubernetes-map-type: atomic
+                      toolModel:
+                        description: Model to use for tool calling, which is less
+                          frequent and often requires more advanced reasoning
+                        type: string
                       url:
                         description: URL is the url this model is queryable on
                         type: string
@@ -236,6 +252,10 @@ spec:
                         - key
                         type: object
                         x-kubernetes-map-type: atomic
+                      toolModel:
+                        description: Model to use for tool calling, which is less
+                          frequent and often requires more advanced reasoning
+                        type: string
                     required:
                     - tokenSecretRef
                     type: object
@@ -288,6 +308,10 @@ spec:
                         - key
                         type: object
                         x-kubernetes-map-type: atomic
+                      toolModel:
+                        description: Model to use for tool calling, which is less
+                          frequent and often requires more advanced reasoning
+                        type: string
                     required:
                     - location
                     - project