22 * Ollama provider implementation using the official Ollama npm package
33 */
44
5- import ollama , { Ollama , ChatResponse , Tool } from 'ollama' ;
5+ import {
6+ ChatRequest as OllamaChatRequest ,
7+ ChatResponse as OllamaChatResponse ,
8+ Ollama ,
9+ ToolCall as OllamaTooCall ,
10+ Tool as OllamaTool ,
11+ Message as OllamaMessage ,
12+ } from 'ollama' ;
13+
614import { TokenUsage } from '../../tokens.js' ;
15+ import { ToolCall } from '../../types.js' ;
716import { LLMProvider } from '../provider.js' ;
817import {
9- FunctionDefinition ,
1018 GenerateOptions ,
1119 LLMResponse ,
1220 Message ,
1321 ProviderOptions ,
22+ FunctionDefinition ,
1423} from '../types.js' ;
1524
1625/**
@@ -31,9 +40,9 @@ export class OllamaProvider implements LLMProvider {
3140
3241 constructor ( model : string , options : OllamaOptions = { } ) {
3342 this . model = model ;
34- const baseUrl =
35- options . baseUrl ||
36- process . env . OLLAMA_BASE_URL ||
43+ const baseUrl =
44+ options . baseUrl ||
45+ process . env . OLLAMA_BASE_URL ||
3746 'http://localhost:11434' ;
3847
3948 this . client = new Ollama ( { host : baseUrl } ) ;
@@ -57,133 +66,165 @@ export class OllamaProvider implements LLMProvider {
5766 // Format messages for Ollama API
5867 const formattedMessages = this . formatMessages ( messages ) ;
5968
60- try {
61- // Prepare chat options
62- const ollamaOptions : Record < string , any > = {
63- temperature,
64- } ;
65-
66- // Add optional parameters if provided
67- if ( topP !== undefined ) ollamaOptions . top_p = topP ;
68- if ( frequencyPenalty !== undefined ) ollamaOptions . frequency_penalty = frequencyPenalty ;
69- if ( presencePenalty !== undefined ) ollamaOptions . presence_penalty = presencePenalty ;
70- if ( maxTokens !== undefined ) ollamaOptions . num_predict = maxTokens ;
71- if ( stopSequences && stopSequences . length > 0 ) ollamaOptions . stop = stopSequences ;
72-
73- // Prepare request parameters
74- const requestParams : any = {
75- model : this . model ,
76- messages : formattedMessages ,
77- stream : false ,
78- options : ollamaOptions ,
69+ // Prepare request options
70+ const requestOptions : OllamaChatRequest = {
71+ model : this . model ,
72+ messages : formattedMessages ,
73+ stream : false ,
74+ options : {
75+ temperature : temperature ,
76+ ...( topP !== undefined && { top_p : topP } ) ,
77+ ...( frequencyPenalty !== undefined && {
78+ frequency_penalty : frequencyPenalty ,
79+ } ) ,
80+ ...( presencePenalty !== undefined && {
81+ presence_penalty : presencePenalty ,
82+ } ) ,
83+ ...( stopSequences &&
84+ stopSequences . length > 0 && { stop : stopSequences } ) ,
85+ } ,
86+ } ;
87+
88+ // Add max_tokens if provided
89+ if ( maxTokens !== undefined ) {
90+ requestOptions . options = {
91+ ...requestOptions . options ,
92+ num_predict : maxTokens ,
7993 } ;
94+ }
8095
81- // Add functions/tools if provided
82- if ( functions && functions . length > 0 ) {
83- requestParams . tools = this . convertFunctionsToTools ( functions ) ;
84- }
96+ // Add functions/tools if provided
97+ if ( functions && functions . length > 0 ) {
98+ requestOptions . tools = this . convertFunctionsToTools ( functions ) ;
99+ }
85100
86- // Make the API request using the Ollama client
87- const response = await this . client . chat ( requestParams ) ;
101+ // Make the API request using the Ollama client
102+ const response : OllamaChatResponse = await this . client . chat ( {
103+ ...requestOptions ,
104+ stream : false ,
105+ } ) ;
88106
89- // Extract content from response
90- const content = response . message ?. content || '' ;
91-
92- // Process tool calls if present
93- const toolCalls = this . processToolCalls ( response ) ;
107+ // Extract content and tool calls
108+ const content = response . message ?. content || '' ;
94109
95- // Create token usage from response data
96- const tokenUsage = new TokenUsage ( ) ;
97- if ( response . prompt_eval_count ) {
98- tokenUsage . input = response . prompt_eval_count ;
99- }
100- if ( response . eval_count ) {
101- tokenUsage . output = response . eval_count ;
102- }
110+ // Handle tool calls if present
111+ const toolCalls = this . extractToolCalls ( response ) ;
103112
104- return {
105- text : content ,
106- toolCalls : toolCalls ,
107- tokenUsage : tokenUsage ,
108- } ;
109- } catch ( error ) {
110- throw new Error ( `Error calling Ollama API: ${ ( error as Error ) . message } ) ;
111- }
113+ // Create token usage from response data
114+ const tokenUsage = new TokenUsage ( ) ;
115+ tokenUsage . output = response . eval_count || 0 ;
116+ tokenUsage . input = response . prompt_eval_count || 0 ;
117+
118+ return {
119+ text : content ,
120+ toolCalls : toolCalls ,
121+ tokenUsage : tokenUsage ,
122+ } ;
112123 }
113124
125+ /*
126+ interface Tool {
127+ type: string;
128+ function: {
129+ name: string;
130+ description: string;
131+ parameters: {
132+ type: string;
133+ required: string[];
134+ properties: {
135+ [key: string]: {
136+ type: string;
137+ description: string;
138+ enum?: string[];
139+ };
140+ };
141+ };
142+ };
143+ }*/
144+
114145 /**
115- * Convert our FunctionDefinition format to Ollama's Tool format
146+ * Convert our function definitions to Ollama tool format
116147 */
117- private convertFunctionsToTools ( functions : FunctionDefinition [ ] ) : Tool [ ] {
118- return functions . map ( ( fn ) => ( {
119- type : 'function' ,
120- function : {
121- name : fn . name ,
122- description : fn . description ,
123- parameters : fn . parameters ,
124- }
125- } ) ) ;
148+ private convertFunctionsToTools (
149+ functions : FunctionDefinition [ ] ,
150+ ) : OllamaTool [ ] {
151+ return functions . map (
152+ ( fn ) =>
153+ ( {
154+ type : 'function' ,
155+ function : {
156+ name : fn . name ,
157+ description : fn . description ,
158+ parameters : fn . parameters ,
159+ } ,
160+ } ) as OllamaTool ,
161+ ) ;
126162 }
127163
128164 /**
129- * Process tool calls from the Ollama response
165+ * Extract tool calls from Ollama response
130166 */
131- private processToolCalls ( response : ChatResponse ) : any [ ] {
132- if ( ! response . message ?. tool_calls || response . message . tool_calls . length === 0 ) {
167+ private extractToolCalls ( response : OllamaChatResponse ) : ToolCall [ ] {
168+ if ( ! response . message ?. tool_calls ) {
133169 return [ ] ;
134170 }
135171
136- return response . message . tool_calls . map ( ( toolCall ) => ( {
137- id : toolCall . function ?. name
138- ? `tool-${ Date . now ( ) } ${ Math . random ( ) . toString ( 36 ) . substring ( 2 , 11 ) }
139- : toolCall . id ,
140- name : toolCall . function ?. name ,
141- content : JSON . stringify ( toolCall . function ?. arguments || { } ) ,
142- } ) ) ;
172+ return response . message . tool_calls . map ( ( toolCall : OllamaTooCall ) => {
173+ //console.log('ollama tool call', toolCall);
174+ return {
175+ id : `tool-${ Date . now ( ) } ${ Math . random ( ) . toString ( 36 ) . substring ( 2 , 11 ) } ,
176+ name : toolCall . function ?. name ,
177+ content :
178+ typeof toolCall . function ?. arguments === 'string'
179+ ? toolCall . function . arguments
180+ : JSON . stringify ( toolCall . function ?. arguments || { } ) ,
181+ } ;
182+ } ) ;
143183 }
144184
145185 /**
146186 * Format messages for Ollama API
147187 */
148- private formatMessages ( messages : Message [ ] ) : any [ ] {
149- return messages . map ( ( msg ) => {
150- if (
151- msg . role === 'user' ||
152- msg . role === 'assistant' ||
153- msg . role === 'system'
154- ) {
155- return {
156- role : msg . role ,
157- content : msg . content ,
158- } ;
159- } else if ( msg . role === 'tool_result' ) {
160- // Ollama expects tool results as a 'tool' role
161- return {
162- role : 'tool' ,
163- content : msg . content ,
164- tool_call_id : msg . tool_use_id ,
165- } ;
166- } else if ( msg . role === 'tool_use' ) {
167- // We'll convert tool_use to assistant messages with tool_calls
168- return {
169- role : 'assistant' ,
170- content : '' ,
171- tool_calls : [
188+ private formatMessages ( messages : Message [ ] ) : OllamaMessage [ ] {
189+ const output : OllamaMessage [ ] = [ ] ;
190+
191+ messages . forEach ( ( msg ) => {
192+ switch ( msg . role ) {
193+ case 'user' :
194+ case 'assistant' :
195+ case 'system' :
196+ output . push ( {
197+ role : msg . role ,
198+ content : msg . content ,
199+ } satisfies OllamaMessage ) ;
200+ break ;
201+ case 'tool_result' :
202+ // Ollama expects tool results as a 'tool' role
203+ output . push ( {
204+ role : 'tool' ,
205+ content :
206+ typeof msg . content === 'string'
207+ ? msg . content
208+ : JSON . stringify ( msg . content ) ,
209+ } as OllamaMessage ) ;
210+ break ;
211+ case 'tool_use' : {
212+ // So there is an issue here is that ollama expects tool calls to be part of the assistant message
213+ // get last message and add tool call to it
214+ const lastMessage : OllamaMessage = output [ output . length - 1 ] ! ;
215+ lastMessage . tool_calls = [
172216 {
173- id : msg . id ,
174217 function : {
175218 name : msg . name ,
176- arguments : msg . content ,
177- }
219+ arguments : JSON . parse ( msg . content ) ,
220+ } ,
178221 } ,
179- ] ,
180- } ;
222+ ] ;
223+ break ;
224+ }
181225 }
182- // Default fallback for unknown message types
183- return {
184- role : 'user' ,
185- content : ( msg as any ) . content || '' ,
186- } ;
187226 } ) ;
227+
228+ return output ;
188229 }
189- }
230+ }
0 commit comments