Add core implementation for AI-powered question answering

Introduce multiple agents, tools, and utilities for processing, extracting, and answering user-provided questions using LLMs and external data. Key features include knowledge processing, question splitting, search term generation, and contextual knowledge handling.

pkg/agents/shared/knowledge.go

@@ -0,0 +1,33 @@
+package shared
+
+import (
+	"strings"
+)
+
+// TidBit is a small piece of information that the AI has learned.
+type TidBit struct {
+	Info   string
+	Source string
+}
+
+type Knowledge struct {
+	// OriginalQuestions are the questions that was asked first to the AI before any processing was done.
+	OriginalQuestions []string
+
+	// RemainingQuestions is the questions that are left to find answers for.
+	RemainingQuestions []string
+
+	// Knowledge are the tidbits of information that the AI has learned.
+	Knowledge []TidBit
+}
+
+// ToMessage converts the knowledge to a message that can be sent to the LLM.
+func (k Knowledge) ToMessage() string {
+	var learned []string
+	for _, t := range k.Knowledge {
+		learned = append(learned, t.Info)
+	}
+	return "Original questions asked:\n" + strings.Join(k.OriginalQuestions, "\n") + "\n" +
+		"Learned information:\n" + strings.Join(learned, "\n") + "\n" +
+		"Remaining questions:\n" + strings.Join(k.RemainingQuestions, "\n")
+}

pkg/agents/shared/modeltracker.go

@@ -0,0 +1,42 @@
+package shared
+
+import (
+	"context"
+	"errors"
+	"sync/atomic"
+
+	gollm "gitea.stevedudenhoeffer.com/steve/go-llm"
+)
+
+type ModelTracker struct {
+	parent  gollm.ChatCompletion
+	maximum int64
+	calls   int64
+}
+
+var _ gollm.ChatCompletion = &ModelTracker{}
+
+// NewModelTracker creates a new model tracker that will limit the number of calls to the parent.
+// Set to 0 to disable the limit.
+func NewModelTracker(parent gollm.ChatCompletion, maximum int64) *ModelTracker {
+	return &ModelTracker{parent: parent, maximum: maximum}
+}
+
+var ErrModelCapacity = errors.New("maximum model capacity reached")
+
+func (m *ModelTracker) ChatComplete(ctx context.Context, req gollm.Request) (gollm.Response, error) {
+	if m.maximum > 0 && atomic.AddInt64(&m.calls, 1) >= m.maximum {
+		return gollm.Response{}, ErrModelCapacity
+	}
+
+	return m.parent.ChatComplete(ctx, req)
+}
+
+// ResetCalls resets the number of calls made to the parent.
+func (m *ModelTracker) ResetCalls() {
+	atomic.StoreInt64(&m.calls, 0)
+}
+
+func (m *ModelTracker) GetCalls() int64 {
+	return atomic.LoadInt64(&m.calls)
+}