refactor(ipuaro): simplify LLM integration with pure XML tool format

Refactor OllamaClient to use pure XML format for tool calls as
designed in CONCEPT.md. Removes dual system (Ollama native tools +
XML parser) in favor of single source of truth (ResponseParser).

Changes:
- Remove tools parameter from ILLMClient.chat() interface
- Remove convertTools(), convertParameters(), extractToolCalls()
- Add XML format instructions to system prompt with examples
- Add CDATA support in ResponseParser for multiline content
- Add tool name validation with helpful error messages
- Move ToolDef/ToolParameter to shared/types/tool-definitions.ts

Benefits:
- Simplified architecture (single source of truth)
- CONCEPT.md compliance (pure XML as designed)
- Better validation (early detection of invalid tools)
- Reduced complexity (fewer format conversions)

Tests: 1444 passed (+4 new tests)
Coverage: 97.83% lines, 91.98% branches, 99.16% functions
This commit is contained in:
imfozilbek
2025-12-01 21:03:55 +05:00
parent 902d1db831
commit 0433ef102c
13 changed files with 290 additions and 212 deletions

View File

@@ -1,15 +1,10 @@
import { type Message, Ollama, type Tool } from "ollama"
import type {
ILLMClient,
LLMResponse,
ToolDef,
ToolParameter,
} from "../../domain/services/ILLMClient.js"
import { type Message, Ollama } from "ollama"
import type { ILLMClient, LLMResponse } from "../../domain/services/ILLMClient.js"
import type { ChatMessage } from "../../domain/value-objects/ChatMessage.js"
import { createToolCall, type ToolCall } from "../../domain/value-objects/ToolCall.js"
import type { LLMConfig } from "../../shared/constants/config.js"
import { IpuaroError } from "../../shared/errors/IpuaroError.js"
import { estimateTokens } from "../../shared/utils/tokens.js"
import { parseToolCalls } from "./ResponseParser.js"
/**
* Ollama LLM client implementation.
@@ -35,19 +30,18 @@ export class OllamaClient implements ILLMClient {
/**
* Send messages to LLM and get response.
* Tool definitions should be included in the system prompt as XML format.
*/
async chat(messages: ChatMessage[], tools?: ToolDef[]): Promise<LLMResponse> {
async chat(messages: ChatMessage[]): Promise<LLMResponse> {
const startTime = Date.now()
this.abortController = new AbortController()
try {
const ollamaMessages = this.convertMessages(messages)
const ollamaTools = tools ? this.convertTools(tools) : undefined
const response = await this.client.chat({
model: this.model,
messages: ollamaMessages,
tools: ollamaTools,
options: {
temperature: this.temperature,
},
@@ -55,15 +49,15 @@ export class OllamaClient implements ILLMClient {
})
const timeMs = Date.now() - startTime
const toolCalls = this.extractToolCalls(response.message)
const parsed = parseToolCalls(response.message.content)
return {
content: response.message.content,
toolCalls,
content: parsed.content,
toolCalls: parsed.toolCalls,
tokens: response.eval_count ?? estimateTokens(response.message.content),
timeMs,
truncated: false,
stopReason: this.determineStopReason(response, toolCalls),
stopReason: this.determineStopReason(response, parsed.toolCalls),
}
} catch (error) {
if (error instanceof Error && error.name === "AbortError") {
@@ -205,69 +199,12 @@ export class OllamaClient implements ILLMClient {
}
}
/**
* Convert ToolDef array to Ollama Tool format.
*/
private convertTools(tools: ToolDef[]): Tool[] {
return tools.map(
(tool): Tool => ({
type: "function",
function: {
name: tool.name,
description: tool.description,
parameters: {
type: "object",
properties: this.convertParameters(tool.parameters),
required: tool.parameters.filter((p) => p.required).map((p) => p.name),
},
},
}),
)
}
/**
* Convert ToolParameter array to JSON Schema properties.
*/
private convertParameters(
params: ToolParameter[],
): Record<string, { type: string; description: string; enum?: string[] }> {
const properties: Record<string, { type: string; description: string; enum?: string[] }> =
{}
for (const param of params) {
properties[param.name] = {
type: param.type,
description: param.description,
...(param.enum && { enum: param.enum }),
}
}
return properties
}
/**
* Extract tool calls from Ollama response message.
*/
private extractToolCalls(message: Message): ToolCall[] {
if (!message.tool_calls || message.tool_calls.length === 0) {
return []
}
return message.tool_calls.map((tc, index) =>
createToolCall(
`call_${String(Date.now())}_${String(index)}`,
tc.function.name,
tc.function.arguments,
),
)
}
/**
* Determine stop reason from response.
*/
private determineStopReason(
response: { done_reason?: string },
toolCalls: ToolCall[],
toolCalls: { name: string; params: Record<string, unknown> }[],
): "end" | "length" | "tool_use" {
if (toolCalls.length > 0) {
return "tool_use"

View File

@@ -27,9 +27,41 @@ const TOOL_CALL_REGEX = /<tool_call\s+name\s*=\s*"([^"]+)">([\s\S]*?)<\/tool_cal
const PARAM_REGEX_NAMED = /<param\s+name\s*=\s*"([^"]+)">([\s\S]*?)<\/param>/gi
const PARAM_REGEX_ELEMENT = /<([a-z_][a-z0-9_]*)>([\s\S]*?)<\/\1>/gi
/**
* CDATA section pattern.
* Matches: <![CDATA[...]]>
*/
const CDATA_REGEX = /<!\[CDATA\[([\s\S]*?)\]\]>/g
/**
* Valid tool names.
* Used for validation to catch typos or hallucinations.
*/
const VALID_TOOL_NAMES = new Set([
"get_lines",
"get_function",
"get_class",
"get_structure",
"edit_lines",
"create_file",
"delete_file",
"find_references",
"find_definition",
"get_dependencies",
"get_dependents",
"get_complexity",
"get_todos",
"git_status",
"git_diff",
"git_commit",
"run_command",
"run_tests",
])
/**
* Parse tool calls from LLM response text.
* Supports XML format: <tool_call name="get_lines"><path>src/index.ts</path></tool_call>
* Validates tool names and provides helpful error messages.
*/
export function parseToolCalls(response: string): ParsedResponse {
const toolCalls: ToolCall[] = []
@@ -41,6 +73,13 @@ export function parseToolCalls(response: string): ParsedResponse {
for (const match of matches) {
const [fullMatch, toolName, paramsXml] = match
if (!VALID_TOOL_NAMES.has(toolName)) {
parseErrors.push(
`Unknown tool "${toolName}". Valid tools: ${[...VALID_TOOL_NAMES].join(", ")}`,
)
continue
}
try {
const params = parseParameters(paramsXml)
const toolCall = createToolCall(
@@ -91,10 +130,16 @@ function parseParameters(xml: string): Record<string, unknown> {
/**
* Parse a value string to appropriate type.
* Supports CDATA sections for multiline content.
*/
function parseValue(value: string): unknown {
const trimmed = value.trim()
const cdataMatches = [...trimmed.matchAll(CDATA_REGEX)]
if (cdataMatches.length > 0 && cdataMatches[0][1] !== undefined) {
return cdataMatches[0][1]
}
if (trimmed === "true") {
return true
}

View File

@@ -23,37 +23,67 @@ export const SYSTEM_PROMPT = `You are ipuaro, a local AI code assistant speciali
3. **Safety**: Confirm destructive operations. Never execute dangerous commands.
4. **Efficiency**: Minimize context usage. Request only necessary code sections.
## Tool Calling Format
When you need to use a tool, format your call as XML:
<tool_call name="tool_name">
<param_name>value</param_name>
<another_param>value</another_param>
</tool_call>
You can call multiple tools in one response. Always wait for tool results before making conclusions.
**Examples:**
<tool_call name="get_lines">
<path>src/index.ts</path>
<start>1</start>
<end>50</end>
</tool_call>
<tool_call name="edit_lines">
<path>src/utils.ts</path>
<start>10</start>
<end>15</end>
<content>const newCode = "hello";</content>
</tool_call>
<tool_call name="find_references">
<symbol>getUserById</symbol>
</tool_call>
## Available Tools
### Reading Tools
- \`get_lines\`: Get specific lines from a file
- \`get_function\`: Get a function by name
- \`get_class\`: Get a class by name
- \`get_structure\`: Get project directory structure
- \`get_lines(path, start?, end?)\`: Get specific lines from a file
- \`get_function(path, name)\`: Get a function by name
- \`get_class(path, name)\`: Get a class by name
- \`get_structure(path?, depth?)\`: Get project directory structure
### Editing Tools (require confirmation)
- \`edit_lines\`: Replace specific lines in a file
- \`create_file\`: Create a new file
- \`delete_file\`: Delete a file
- \`edit_lines(path, start, end, content)\`: Replace specific lines in a file
- \`create_file(path, content)\`: Create a new file
- \`delete_file(path)\`: Delete a file
### Search Tools
- \`find_references\`: Find all usages of a symbol
- \`find_definition\`: Find where a symbol is defined
- \`find_references(symbol, path?)\`: Find all usages of a symbol
- \`find_definition(symbol)\`: Find where a symbol is defined
### Analysis Tools
- \`get_dependencies\`: Get files this file imports
- \`get_dependents\`: Get files that import this file
- \`get_complexity\`: Get complexity metrics
- \`get_todos\`: Find TODO/FIXME comments
- \`get_dependencies(path)\`: Get files this file imports
- \`get_dependents(path)\`: Get files that import this file
- \`get_complexity(path?, limit?)\`: Get complexity metrics
- \`get_todos(path?, type?)\`: Find TODO/FIXME comments
### Git Tools
- \`git_status\`: Get repository status
- \`git_diff\`: Get uncommitted changes
- \`git_commit\`: Create a commit (requires confirmation)
- \`git_status()\`: Get repository status
- \`git_diff(path?, staged?)\`: Get uncommitted changes
- \`git_commit(message, files?)\`: Create a commit (requires confirmation)
### Run Tools
- \`run_command\`: Execute a shell command (security checked)
- \`run_tests\`: Run the test suite
- \`run_command(command, timeout?)\`: Execute a shell command (security checked)
- \`run_tests(path?, filter?, watch?)\`: Run the test suite
## Response Guidelines

View File

@@ -1,4 +1,4 @@
import type { ToolDef } from "../../domain/services/ILLMClient.js"
import type { ToolDef } from "../../shared/types/tool-definitions.js"
/**
* Tool definitions for ipuaro LLM.