Unverified Commit e8b981fa authored by Parth Sareen's avatar Parth Sareen Committed by GitHub
Browse files

tools: refactor tool call parsing and enable streaming (#10415)

parent 884d2609
package tools
import (
"encoding/json"
"errors"
"log/slog"
"strings"
gotmpl "text/template"
"github.com/ollama/ollama/api"
"github.com/ollama/ollama/template"
)
var (
errInvalidToolCall = errors.New("invalid tool call format")
errAccumulateMore = errors.New("need to accumulate more content")
)
type Parser struct {
parseLeadingJSON bool
prefix string
prefixFound bool
tmpl gotmpl.Template
sb strings.Builder
index int
name string
arguments string
done bool
}
// parseJSONToolCalls attempts to parse a JSON string into a slice of ToolCalls.
//
// Parameters:
// - s: The string to parse
// - name: The field name from template that identifies the tool call name
// - arguments: The field name from template that identifies the tool call arguments
//
// Returns:
// - []api.ToolCall: The parsed tool calls if successful
// - error: ErrAccumulateMore if braces unbalanced, ErrInvalidToolCall if invalid, or nil if successful
func parseJSONToolCalls(s string, name, arguments string, prefix string) ([]api.ToolCall, error) {
// Check for balanced braces before attempting to parse
braceCount := 0
squareCount := 0
startIndex := -1
var rawToolCalls []string
s = strings.TrimSpace(s)
// Only track these if we don't have a prefix as it will be cut off from the prefix. Also track in the parseLeadingJSON case.
trackSquareBrackets := prefix == "" || !strings.HasSuffix(prefix, "[") || strings.HasPrefix(s, "[")
for i, c := range s {
switch c {
case '{':
braceCount++
if startIndex == -1 {
startIndex = i
}
case '}':
braceCount--
if braceCount == 0 {
rawToolCalls = append(rawToolCalls, s[startIndex:i+1])
startIndex = -1
}
case '[':
if trackSquareBrackets {
squareCount++
}
case ']':
if trackSquareBrackets {
squareCount--
}
}
// Negative means we have an extra closing brace/bracket
if braceCount < 0 || squareCount < 0 {
return nil, errInvalidToolCall
}
}
// If braces/brackets aren't balanced, need more input
if braceCount > 0 || squareCount > 0 {
return nil, errAccumulateMore
}
t := strings.TrimSpace(s)
if len(t) == 0 {
return nil, errAccumulateMore
}
// If the input is a single square bracket, it's not a valid tool call
if t[0] == '[' && len(t) == 1 {
return nil, errAccumulateMore
}
// Attempt full unmarshal of the JSON
var toolCalls []api.ToolCall
for _, rawToolCall := range rawToolCalls {
var resp map[string]any
if err := json.Unmarshal([]byte(rawToolCall), &resp); err != nil {
continue
}
// Collect nested objects that could contain tool calls
objs := collect(resp)
if len(objs) == 0 {
continue
}
// Extract tool calls from objects
for _, kv := range objs {
n, nok := kv[name].(string)
a, aok := kv[arguments].(map[string]any)
if nok && aok {
toolCalls = append(toolCalls, api.ToolCall{
Function: api.ToolCallFunction{
Name: n,
Arguments: a,
},
})
} else {
slog.Debug("No valid tool call found in object.", "object", kv)
}
}
}
// Valid JSON, no tool calls found
if len(toolCalls) == 0 {
slog.Debug("No valid tool calls found in any raw tool calls.", "rawToolCalls", rawToolCalls)
return nil, errInvalidToolCall
}
return toolCalls, nil
}
// checkPrefix processes a string to find and handle a prefix pattern.
//
// Returns:
// - The processed string with prefix removed if found
// - error: ErrAccumulateMore if prefix is incomplete, or nil if successful
func (p *Parser) checkPrefix(s string) (string, error) {
original := s
if strings.ContainsRune(s, '\n') {
s = strings.ReplaceAll(s, "\n", " ")
}
if s == "" || p.prefix == "" {
return s, nil
}
// Check for prefix at start of string
if cut, hasPrefix := strings.CutPrefix(s, p.prefix); hasPrefix {
// Found prefix at start - accumulate for potential tool
p.prefixFound = true
return cut, nil
}
// Check if prefix overlaps end of string
if idx := suffixOverlap(s, p.prefix); idx != -1 {
// Return everything except overlapping portion
p.sb.Reset()
p.sb.WriteString(s[idx:])
return original[:idx], errAccumulateMore
}
// Check if prefix appears in middle of string
if idx := strings.Index(s, p.prefix); idx != -1 {
// Save remainder starting at prefix for next pass
p.sb.Reset()
p.sb.WriteString(strings.TrimSpace(s[idx:]))
// Return everything before prefix
return original[:idx], errAccumulateMore
}
// No partial prefix found
return s, nil
}
// Add processes a string input to parse tool calls and content.
// It handles prefix detection and JSON parsing to extract tool calls.
//
// Returns:
// - tools: Any parsed tool calls
// - content: Non-tool call content
func (p *Parser) Add(s string) (tools []api.ToolCall, content string) {
if strings.TrimSpace(s) == "" {
return nil, s
}
if p.done {
if p.index == 0 {
// Return original string if no tool calls found at start
return nil, s
}
// Return empty if no tool calls found after start
return nil, ""
}
p.sb.WriteString(s)
s = p.sb.String()
// Check for prefix pattern in input
s, err := p.checkPrefix(s)
if err != nil {
// Need more input to complete prefix
return nil, s
}
// Exit if prefix exists in template, greedy parsing is off, and prefix not found
if !p.parseLeadingJSON && !p.prefixFound {
p.sb.Reset()
return nil, s
}
toolCalls, err := parseJSONToolCalls(s, p.name, p.arguments, p.prefix)
if err != nil {
if errors.Is(err, errAccumulateMore) {
return nil, ""
}
p.sb.Reset()
// Do not try parsing leading JSON if JSON not found
p.parseLeadingJSON = false
if p.prefix == "" {
p.done = true
}
if p.index != 0 && p.prefix == "" {
return nil, ""
}
if p.prefixFound {
// Drop tokens since prefix was found
return nil, ""
}
return nil, s
}
for _, tc := range toolCalls {
tc.Function.Index = p.index
p.index++
}
p.sb.Reset()
return toolCalls, ""
}
// NewParser creates a new tool call parser from a template. It extracts the tool call format,
// prefix, and field names from the template to use for parsing tool calls from model output.
//
// Returns an error if the template does not contain valid tool call formatting.
func NewParser(templateToProcess *gotmpl.Template) (*Parser, error) {
parsed, err := template.Parse(templateToProcess.Root.String())
if err != nil {
return nil, err
}
tt, err := toolTemplate(parsed)
if err != nil {
return nil, err
}
tp := toolPrefix(templateToProcess)
name, arguments, err := extractToolArgs(tt)
if err != nil {
return nil, err
}
return &Parser{
tmpl: *tt,
sb: strings.Builder{},
prefix: tp,
parseLeadingJSON: true,
name: name,
arguments: arguments,
}, nil
}
This diff is collapsed.
package tools
import (
"bytes"
"encoding/json"
"errors"
"log/slog"
"slices"
"strings"
gotmpl "text/template"
"text/template/parse"
"github.com/ollama/ollama/api"
"github.com/ollama/ollama/template"
)
// extractToolCallsFormat traverses a template AST to find text that follows a ".ToolCalls" condition.
// It walks the template nodes looking for if-statements containing ".ToolCalls" and extracts any
// immediate text nodes that follow. This is used to identify tool call prefixes and formatting.
//
// Returns:
// - string: The extracted text following the first ".ToolCalls" condition found
// - bool: Whether a ".ToolCalls" condition was found in the template
func extractToolCallsFormat(tmpl *gotmpl.Template) (string, bool) {
if tmpl == nil || tmpl.Tree == nil {
slog.Debug("template or tree is nil")
return "", false
}
var result string
var found bool
var walk func(nodes []parse.Node)
walk = func(nodes []parse.Node) {
for _, node := range nodes {
if found {
return
}
switch n := node.(type) {
case *parse.IfNode:
if isToolCallsNode(n) {
// Collect immediate TextNode(s) at start of IfNode's list
var sb strings.Builder
for _, innerNode := range n.List.Nodes {
if tn, ok := innerNode.(*parse.TextNode); ok {
sb.Write(tn.Text)
} else {
// Stop at first non-text node
break
}
}
result = sb.String()
found = true
return
}
// Recurse into child nodes
walk(n.List.Nodes)
if n.ElseList != nil {
walk(n.ElseList.Nodes)
}
case *parse.ListNode:
walk(n.Nodes)
case *parse.RangeNode:
walk(n.List.Nodes)
if n.ElseList != nil {
walk(n.ElseList.Nodes)
}
case *parse.WithNode:
walk(n.List.Nodes)
if n.ElseList != nil {
walk(n.ElseList.Nodes)
}
default:
// Continue to next node
continue
}
}
}
walk(tmpl.Tree.Root.Nodes)
return result, found
}
// isToolCallsNode detects if a node's condition includes ".ToolCalls"
func isToolCallsNode(n *parse.IfNode) bool {
for _, cmd := range n.Pipe.Cmds {
for _, arg := range cmd.Args {
if field, ok := arg.(*parse.FieldNode); ok {
if slices.Contains(field.Ident, "ToolCalls") {
return true
}
}
}
}
return false
}
func toolPrefix(tmpl *gotmpl.Template) string {
tokenText, ok := extractToolCallsFormat(tmpl)
if !ok {
return ""
}
tokenText = strings.TrimSpace(tokenText)
tokenText = strings.ReplaceAll(tokenText, "\r", "")
tokenText = strings.ReplaceAll(tokenText, "\n", " ")
return tokenText
}
// toolTemplate creates a subtree from the node that ranges over .ToolCalls
//
// Returns:
// - *gotmpl.Template: The subtree containing the .ToolCalls range
// - error: Error if parsing failed
func toolTemplate(t *template.Template) (*gotmpl.Template, error) {
tmpl := t.Subtree(func(n parse.Node) bool {
if t, ok := n.(*parse.RangeNode); ok {
return slices.Contains(template.Identifiers(t.Pipe), "ToolCalls")
}
return false
})
if tmpl == nil {
return nil, errors.New("failed to find tool template")
}
return tmpl, nil
}
// suffixOverlap returns the index in s where the longest suffix overlap with prefix begins
//
// Returns:
// - int: The starting index in s where the suffix overlap begins
func suffixOverlap(s, prefix string) int {
max := min(len(prefix), len(s))
for i := max; i > 0; i-- {
if strings.HasSuffix(s, prefix[:i]) {
return len(s) - i
}
}
return -1
}
// extractToolArgs executes a template with a known tool call format to extract the name and arguments
//
// Returns:
// - string: The name of the tool call
// - string: The arguments of the tool call
// - error: Error if parsing failed
func extractToolArgs(tmpl *gotmpl.Template) (name, arguments string, err error) {
var b bytes.Buffer
if err := tmpl.Execute(&b, map[string][]api.ToolCall{
"ToolCalls": {
{
Function: api.ToolCallFunction{
Name: "@@name@@",
Arguments: api.ToolCallFunctionArguments{
"@@argument@@": 1,
},
},
},
},
}); err != nil {
return "", "", err
}
var obj any
err = json.Unmarshal(b.Bytes(), &obj)
if err != nil {
return "", "", err
}
var objs []map[string]any
switch v := obj.(type) {
case map[string]any:
objs = []map[string]any{v}
case []map[string]any:
objs = v
case []any:
objs = collect(v)
}
if len(objs) == 0 {
return "", "", errors.New("no template objects found")
}
// find the keys that correspond to the name and arguments fields
for k, v := range objs[0] {
switch v.(type) {
case string:
name = k
case map[string]any:
arguments = k
}
}
if name == "" || arguments == "" {
slog.Debug("missing required fields in tool call template", "name", name, "arguments", arguments)
return "", "", errors.New("missing required fields in tool call template")
}
return name, arguments, nil
}
// collect recursively traverses an object to collect all nested maps
//
// Returns:
// - []map[string]any: A slice of all nested maps found in the object
func collect(obj any) []map[string]any {
var all []map[string]any
switch o := obj.(type) {
case map[string]any:
all = append(all, o)
for _, v := range o {
all = append(all, collect(v)...)
}
case []any:
for _, v := range o {
all = append(all, collect(v)...)
}
default:
return nil
}
return all
}
package tools
import (
"testing"
gotmpl "text/template"
"github.com/ollama/ollama/template"
)
func TestExtractToolCallsFormat(t *testing.T) {
cases := []struct {
name string
template string
want string
found bool
}{
{
name: "nil template",
template: "",
want: "",
found: false,
},
{
name: "basic tool call with text",
template: "{{if .ToolCalls}}Hello world{{end}}",
want: "Hello world",
found: true,
},
{
name: "tool call with json format",
template: "{{if .ToolCalls}}```json\n{{end}}",
want: "```json\n",
found: true,
},
{
name: "tool call in range",
template: "{{range .ToolCalls}}tool: {{.}}{{end}}",
want: "",
found: false,
},
{
name: "tool call with multiple text nodes",
template: "{{if .ToolCalls}}First text{{if .Something}}inner{{end}}Second text{{end}}",
want: "First text",
found: true,
},
{
name: "nested if without tool calls",
template: "{{if .Something}}{{if .OtherThing}}text{{end}}{{end}}",
want: "",
found: false,
},
}
for _, tc := range cases {
t.Run(tc.name, func(t *testing.T) {
tmpl, err := gotmpl.New("test").Parse(tc.template)
if err != nil && tc.template != "" {
t.Fatalf("failed to parse template: %v", err)
}
got, found := extractToolCallsFormat(tmpl)
if got != tc.want {
t.Errorf("got text %q, want %q", got, tc.want)
}
if found != tc.found {
t.Errorf("got found %v, want %v", found, tc.found)
}
})
}
}
func TestToolPrefix(t *testing.T) {
cases := []struct {
name string
template string
want string
}{
{
name: "basic tool call with action prefix",
template: "{{if .ToolCalls}}Action: ```json{{end}}",
want: "Action: ```json",
},
{
name: "incomplete functools bracket",
template: "{{if .ToolCalls}}functools[{{end}}",
want: "functools[",
},
{
name: "tool call with angle brackets",
template: "{{if .ToolCalls}}Hello, world! <tool_call>{{end}}",
want: "Hello, world! <tool_call>",
},
{
name: "multiple tool call formats",
template: "{{if .ToolCalls}}[tool_call] <tool_call>{{end}}",
want: "[tool_call] <tool_call>",
},
{
name: "single angle bracket tool call",
template: "{{if .ToolCalls}}<tool_call>{{end}}",
want: "<tool_call>",
},
{
name: "incomplete angle bracket after tool call",
template: "{{if .ToolCalls}}[tool_call] <{{end}}",
want: "[tool_call] <",
},
{
name: "angle bracket prefix with tool call",
template: "{{if .ToolCalls}}> <tool_call>{{end}}",
want: "> <tool_call>",
},
{
name: "uppercase tool call with incomplete bracket",
template: "{{if .ToolCalls}}[TOOL_CALL] [{{end}}",
want: "[TOOL_CALL] [",
},
{
name: "uppercase tool call with adjacent bracket",
template: "{{if .ToolCalls}}[TOOL_CALL][{{end}}",
want: "[TOOL_CALL][",
},
{
name: "tool call with pipe delimiters",
template: "{{if .ToolCalls}}<|tool_call|>{{end}}",
want: "<|tool_call|>",
},
{
name: "tool with no prefix",
template: "{{if .ToolCalls}}{{end}}",
want: "",
},
}
for _, tt := range cases {
t.Run(tt.name, func(t *testing.T) {
tmpl, err := gotmpl.New("test").Parse(tt.template)
if err != nil {
t.Fatalf("failed to parse template: %v", err)
}
got := toolPrefix(tmpl)
if got != tt.want {
t.Errorf("ToolToken(%q) = %q; want %q", tt.template, got, tt.want)
}
})
}
}
func TestToolTemplate(t *testing.T) {
cases := []struct {
name string
template string
want bool
}{
{
name: "basic tool call range",
template: "{{range .ToolCalls}}test{{end}}",
want: true,
},
{
name: "no tool calls",
template: "{{range .Other}}test{{end}}",
want: false,
},
{
name: "nested tool calls",
template: "{{range .Outer}}{{range .ToolCalls}}test{{end}}{{end}}",
want: true,
},
{
name: "empty template",
template: "",
want: false,
},
{
name: "tool calls in if statement",
template: "{{if .ToolCalls}}test{{end}}",
want: false,
},
}
for _, tt := range cases {
t.Run(tt.name, func(t *testing.T) {
tmpl, err := gotmpl.New("test").Parse(tt.template)
if err != nil {
t.Fatalf("failed to parse template: %v", err)
}
parsed, err := template.Parse(tmpl.Root.String())
if err != nil {
t.Fatalf("failed to parse template: %v", err)
}
_, err = toolTemplate(parsed)
if err != nil && tt.want {
t.Errorf("toolTemplate() = %v; want %v", err, tt.want)
}
})
}
}
func TestSuffixOverlap(t *testing.T) {
cases := []struct {
name string
s string
d string
want int
}{
{
name: "no overlap",
s: "hello world",
d: "<tool_call>",
want: -1,
},
{
name: "full overlap",
s: "<tool_call>",
d: "<tool_call>",
want: 0,
},
{
name: "partial overlap",
s: "text <tool_call>",
d: "<tool_call>",
want: 5,
},
{
name: "delimiter longer than string",
s: "<tool>",
d: "<tool_call>",
want: -1,
},
{
name: "empty string",
s: "",
d: "<tool_call>",
want: -1,
},
{
name: "empty delimiter",
s: "<tool_call>",
d: "",
want: -1,
},
{
name: "single char overlap",
s: "test<",
d: "<tool_call>",
want: 4,
},
{
name: "partial tool call",
s: "hello <tool_",
d: "<tool_call>",
want: 6,
},
}
for _, tt := range cases {
t.Run(tt.name, func(t *testing.T) {
got := suffixOverlap(tt.s, tt.d)
if got != tt.want {
t.Errorf("suffixOverlap(%q, %q) = %d; want %d", tt.s, tt.d, got, tt.want)
}
})
}
}
func TestExtractToolArgs(t *testing.T) {
cases := []struct {
name string
template string
want string
ok bool
}{
{
name: "basic tool call with text after",
template: `{{if .ToolCalls}}tool response{{end}}`,
want: "tool response",
ok: true,
},
{
name: "tool call with mixed content after",
template: `{{if .ToolCalls}}<tool_call>{{.Something}}{{end}}`,
want: "<tool_call>",
ok: true,
},
{
name: "tool call with no text after",
template: `{{if .ToolCalls}}{{.Something}}{{end}}`,
want: "",
ok: true,
},
{
name: "nested tool call",
template: `{{if .Something}}{{if .ToolCalls}}[TOOL_CALL]{{end}}{{end}}`,
want: "[TOOL_CALL]",
ok: true,
},
{
name: "no tool calls",
template: `{{if .Something}}no tools here{{end}}`,
want: "",
ok: false,
},
{
name: "empty template",
template: ``,
want: "",
ok: false,
},
{
name: "multiple tool calls sections",
template: `{{if .ToolCalls}}first{{end}}{{if .ToolCalls}}second{{end}}`,
want: "first",
ok: true,
},
{
name: "range over tool calls",
template: `{{if .ToolCalls}}{{range .ToolCalls}}tool{{end}}{{end}}`,
want: "",
ok: true,
},
{
name: "tool calls with pipe delimiters",
template: `{{if .ToolCalls}}<|tool|>{{end}}`,
want: "<|tool|>",
ok: true,
},
{
name: "tool calls with nested template",
template: `{{if .ToolCalls}}{{template "tool" .}}{{end}}`,
want: "",
ok: true,
},
{
name: "tool calls with whitespace variations",
template: `{{if .ToolCalls}} tool {{end}}`,
want: " tool ",
ok: true,
},
}
for _, tt := range cases {
t.Run(tt.name, func(t *testing.T) {
tmpl, err := gotmpl.New("test").Parse(tt.template)
if err != nil {
t.Fatalf("failed to parse template: %v", err)
}
got, ok := extractToolCallsFormat(tmpl)
if got != tt.want {
t.Errorf("TextAfterToolCalls() got = %q, want %q", got, tt.want)
}
if ok != tt.ok {
t.Errorf("TextAfterToolCalls() ok = %v, want %v", ok, tt.ok)
}
})
}
}
func TestCollect(t *testing.T) {
cases := []struct {
name string
obj any
want []map[string]any
}{
{
name: "simple map",
obj: map[string]any{
"key": "value",
},
want: []map[string]any{
{"key": "value"},
},
},
{
name: "nested map",
obj: map[string]any{
"outer": map[string]any{
"inner": "value",
},
},
want: []map[string]any{
{"outer": map[string]any{"inner": "value"}},
{"inner": "value"},
},
},
{
name: "array of maps",
obj: []any{
map[string]any{"key1": "val1"},
map[string]any{"key2": "val2"},
},
want: []map[string]any{
{"key1": "val1"},
{"key2": "val2"},
},
},
{
name: "deeply nested",
obj: map[string]any{
"l1": map[string]any{
"l2": map[string]any{
"l3": "value",
},
},
},
want: []map[string]any{
{"l1": map[string]any{"l2": map[string]any{"l3": "value"}}},
{"l2": map[string]any{"l3": "value"}},
{"l3": "value"},
},
},
{
name: "non-map value",
obj: "string",
want: nil,
},
}
for _, tt := range cases {
t.Run(tt.name, func(t *testing.T) {
got := collect(tt.obj)
if len(got) != len(tt.want) {
t.Errorf("collect() got %d maps, want %d", len(got), len(tt.want))
return
}
// Compare each map in the result
for i := range tt.want {
if !mapsEqual(got[i], tt.want[i]) {
t.Errorf("collect() map[%d] = %v, want %v", i, got[i], tt.want[i])
}
}
})
}
}
// mapsEqual compares two maps for deep equality
func mapsEqual(m1, m2 map[string]any) bool {
if len(m1) != len(m2) {
return false
}
for k, v1 := range m1 {
v2, ok := m2[k]
if !ok {
return false
}
switch val1 := v1.(type) {
case map[string]any:
val2, ok := v2.(map[string]any)
if !ok || !mapsEqual(val1, val2) {
return false
}
default:
if v1 != v2 {
return false
}
}
}
return true
}
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment