Skip to content

RAGFlow is an agent-native database: a Retrieval-Augmented Generation (RAG) engine that grew into context infrastructure.

18.3 KB
Raw Blame
//
//  Copyright 2026 The InfiniFlow Authors. All Rights Reserved.
//
//  Licensed under the Apache License, Version 2.0 (the "License");
//  you may not use this file except in compliance with the License.
//  You may obtain a copy of the License at
//
//      http://www.apache.org/licenses/LICENSE-2.0
//
//  Unless required by applicable law or agreed to in writing, software
//  distributed under the License is distributed on an "AS IS" BASIS,
//  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
//  See the License for the specific language governing permissions and
//  limitations under the License.
//

package models

import (
	"context"
	"encoding/json"
	"fmt"
	"strings"
	"sync"

	"ragflow/internal/tokenizer"
)

// recordCallUsage feeds ONE chat call's token split to the context-level run sink (if
// installed). The caller passes the numbers it already holds: the ChatModel is shared
// by every call on it, so a per-call field would let a concurrent call's tokens be
// attributed to this one. The canvas component gets its own copy from the message
// metadata (ResponseMeta.Usage) for the same reason.
func recordCallUsage(ctx context.Context, cm *ChatModel, prompt, completion, total int) {
	if cm == nil {
		return
	}
	model := ""
	if cm.ModelName != nil {
		model = *cm.ModelName
	}
	recordUsage(ctx, model, &TokenUsage{PromptTokens: prompt, CompletionTokens: completion, TotalTokens: total})
}

// recordUsage records one chat call's token usage to the context-level run sink (if
// installed). The caller passes the usage its own call returned: nothing about one
// call's tokens is parked on the shared *ChatModel, where a concurrent call could
// replace them between the write and the read and attribute them to the wrong request.
func recordUsage(ctx context.Context, model string, usage *TokenUsage) {
	if usage == nil {
		return
	}
	tokenizer.RecordRunTokenUsageFor(ctx, model, usage.PromptTokens, usage.CompletionTokens, usage.TotalTokens)
}

const (
	defaultMaxRetries = 3
	defaultMaxRounds  = 5
)

// ChatWithTools runs the non-streaming tool-calling loop.
func (cm *ChatModel) ChatWithTools(ctx context.Context, system string, history []Message, chatCfg *ChatConfig) (string, int, error) {
	tc := cm.ToolConfig
	if tc == nil {
		return "", 0, fmt.Errorf("ChatWithTools called without bound tools")
	}

	var toolsList interface{}
	if err := json.Unmarshal([]byte(tc.Tools), &toolsList); err != nil {
		return "", 0, fmt.Errorf("failed to parse tools JSON: %w", err)
	}

	maxRounds := tc.MaxRounds
	if maxRounds <= 0 {
		maxRounds = defaultMaxRounds
	}
	maxRetries := tc.MaxRetries
	if maxRetries <= 0 {
		maxRetries = defaultMaxRetries
	}

	if system != "" && len(history) > 0 && history[0].Role != "system" {
		history = append([]Message{{Role: "system", Content: system}}, history...)
	}

	baseHistory := make([]Message, len(history))
	copy(baseHistory, history)

	for attempt := 0; attempt < maxRetries; attempt++ {
		select {
		case <-ctx.Done():
			return "", 0, ctx.Err()
		default:
		}

		h := make([]Message, len(baseHistory))
		copy(h, baseHistory)

		answer, tokens, err := runToolLoop(ctx, cm, h, toolsList, chatCfg, maxRounds)
		if err == nil {
			return answer, tokens, nil
		}
	}
	return "", 0, fmt.Errorf("ChatWithTools failed after %d retries", maxRetries)
}

func runToolLoop(ctx context.Context, cm *ChatModel, history []Message, toolsList interface{}, chatCfg *ChatConfig, maxRounds int) (string, int, error) {
	// Aggregate prompt/completion/total across all tool-calling rounds.
	// Mirrors Python PR #16420 fix: previously the total was overwritten
	// each round; now we accumulate so multi-round tool conversations
	// report the correct grand total.
	var totalTokens int
	aggUsage := &TokenUsage{}

	addRoundUsage := func(resp *ChatResponse) {
		u := resp.Usage
		if u == nil {
			return
		}
		aggUsage.PromptTokens += u.PromptTokens
		aggUsage.CompletionTokens += u.CompletionTokens
		aggUsage.TotalTokens += u.TotalTokens
		totalTokens = aggUsage.TotalTokens
		// Store per-round delta (not cumulative) so RecordRunTokenUsage
		// records each round's contribution exactly once.
		recordCallUsage(ctx, cm, u.PromptTokens, u.CompletionTokens, u.TotalTokens)
	}

	for round := 0; round <= maxRounds; round++ {
		select {
		case <-ctx.Done():
			return "", totalTokens, ctx.Err()
		default:
		}
		cfg := *chatCfg
		cfg.Tools = toolsList
		tcChoice := "auto"
		cfg.ToolChoice = &tcChoice

		resp, err := cm.ModelDriver.ChatWithMessages(ctx, *cm.ModelName, history, cm.APIConfig, &cfg, nil)
		if err != nil {
			return "", totalTokens, fmt.Errorf("round %d: %w", round, err)
		}
		if resp == nil {
			return "", totalTokens, fmt.Errorf("round %d: nil response", round)
		}
		addRoundUsage(resp)

		if len(resp.ToolCalls) == 0 {
			answer := ""
			if resp.Answer != nil {
				answer = *resp.Answer
			}
			if resp.ReasonContent != nil && *resp.ReasonContent != "" {
				answer = "<think>" + *resp.ReasonContent + "</think>" + answer
			}
			// Fallback: if the provider didn't return usage info,
			// estimate from the answer text using tiktoken.
			if resp.Usage == nil || resp.Usage.TotalTokens == 0 {
				totalTokens += tokenizer.NumTokensFromString(answer)
			}
			return answer, totalTokens, nil
		}

		// Execute the round's tool calls and fold their results into history.
		// If one of them is a terminal tool and succeeded, its result is already
		// the final answer — return it directly instead of re-invoking the model
		// (Python chat_model.py:619-627). Non-streaming matches Python's
		// all-empty fallthrough: no qualifying terminal result and the loop
		// runs another round (chat_model.py:692-704).
		var hit bool
		var toolAnswer string
		history, toolAnswer, hit = appendToolResults(history, resp.ToolCalls, cm.ToolConfig.ToolCallSession, terminalSet(cm), false)
		if hit {
			// This round's usage was already folded in by addRoundUsage above
			// (it accumulates into aggUsage and forwards a per-round delta to the
			// run-usage sink, which ADDS): calling it again here would count the
			// round twice in both totalTokens and the recorded run usage.
			return toolAnswer, totalTokens, nil
		}
		// history now carries this round's tool results; continue to the next round.
	}

	// Exceeded max rounds
	history = append(history, Message{
		Role:    "user",
		Content: fmt.Sprintf("Exceed max rounds: %d", maxRounds),
	})
	cfg := *chatCfg
	resp, err := cm.ModelDriver.ChatWithMessages(ctx, *cm.ModelName, history, cm.APIConfig, &cfg, nil)
	if err != nil {
		return "", totalTokens, fmt.Errorf("final call: %w", err)
	}
	if resp == nil || resp.Answer == nil {
		return "", totalTokens, fmt.Errorf("final call: no answer")
	}
	addRoundUsage(resp)
	// Fallback: use text-based estimation if no authoritative usage.
	if resp.Usage == nil || resp.Usage.TotalTokens == 0 {
		totalTokens += tokenizer.NumTokensFromString(*resp.Answer)
	}
	return *resp.Answer, totalTokens, nil
}

// terminalSet returns the configured terminal-tool name set, or nil when none.
func terminalSet(cm *ChatModel) map[string]struct{} {
	if cm.ToolConfig == nil {
		return nil
	}
	return cm.ToolConfig.TerminalTools
}

// ChatStreamlyWithTools runs the streaming tool-calling loop.
func (cm *ChatModel) ChatStreamlyWithTools(ctx context.Context, system string, history []Message, chatCfg *ChatConfig, sender func(*string, *string) error) (int, error) {
	tc := cm.ToolConfig
	if tc == nil {
		return 0, fmt.Errorf("ChatStreamlyWithTools called without bound tools")
	}

	var toolsList interface{}
	if err := json.Unmarshal([]byte(tc.Tools), &toolsList); err != nil {
		return 0, fmt.Errorf("failed to parse tools JSON: %w", err)
	}

	maxRounds := tc.MaxRounds
	if maxRounds <= 0 {
		maxRounds = defaultMaxRounds
	}
	maxRetries := tc.MaxRetries
	if maxRetries <= 0 {
		maxRetries = defaultMaxRetries
	}

	if system != "" && len(history) > 0 && history[0].Role != "system" {
		history = append([]Message{{Role: "system", Content: system}}, history...)
	}

	baseHistory := make([]Message, len(history))
	copy(baseHistory, history)

	for attempt := 0; attempt < maxRetries; attempt++ {
		select {
		case <-ctx.Done():
			return 0, ctx.Err()
		default:
		}

		h := make([]Message, len(baseHistory))
		copy(h, baseHistory)

		totalTokens, err := runStreamToolLoop(ctx, cm, h, toolsList, chatCfg, maxRounds, sender)
		if err == nil {
			return totalTokens, nil
		}
	}
	return 0, fmt.Errorf("ChatStreamlyWithTools failed after %d retries", maxRetries)
}

func runStreamToolLoop(ctx context.Context, cm *ChatModel, history []Message, toolsList interface{}, chatCfg *ChatConfig, maxRounds int, sender func(*string, *string) error) (int, error) {
	// Aggregate token counts across every tool-calling round (each round is a
	// separate provider request). Committing per round avoids the previous
	// bug where a later round's total overwrote earlier rounds.
	var totalTokens int
	aggUsage := &TokenUsage{}

	commitRound := func(cfg *ChatConfig, roundTokens int) {
		// Prefer the authoritative usage from the API (extracted via
		// stream_options.include_usage=true) over text-based token
		// counting. Mirrors Python's usage_from_response accumulation
		// in chat_model.py streaming handlers.
		// Track per-round delta so RecordRunTokenUsage records each
		// round's contribution exactly once (the sink does Add, not Set).
		var deltaPrompt, deltaCompletion, deltaTotal int
		if u := cfg.UsageResult; u != nil && u.TotalTokens > 0 {
			deltaPrompt, deltaCompletion, deltaTotal = u.PromptTokens, u.CompletionTokens, u.TotalTokens
			aggUsage.PromptTokens += deltaPrompt
			aggUsage.CompletionTokens += deltaCompletion
			aggUsage.TotalTokens += deltaTotal
		} else {
			deltaTotal = roundTokens
			aggUsage.TotalTokens += deltaTotal
		}
		totalTokens = aggUsage.TotalTokens
		recordCallUsage(ctx, cm, deltaPrompt, deltaCompletion, deltaTotal)
	}

	for round := 0; round <= maxRounds; round++ {
		select {
		case <-ctx.Done():
			return totalTokens, ctx.Err()
		default:
		}
		cfg := *chatCfg
		cfg.Tools = toolsList
		tcChoice := "auto"
		cfg.ToolChoice = &tcChoice
		cfg.Stream = boolPtr(true)
		var tcs []map[string]interface{}
		cfg.ToolCallsResult = &tcs
		var roundUsage TokenUsage
		cfg.UsageResult = &roundUsage

		reasoningStarted := false
		var answer string
		var pendingThinkClose bool
		var roundTokens int

		err := cm.ModelDriver.ChatStreamlyWithSender(ctx, *cm.ModelName, history, cm.APIConfig, &cfg, nil, func(delta *string, reason *string) error {
			if reason != nil && *reason != "" {
				if !reasoningStarted {
					reasoningStarted = true
					thinkOpen := "<think>"
					if e := sender(&thinkOpen, nil); e != nil {
						return e
					}
				}
				pendingThinkClose = true
				roundTokens += tokenizer.NumTokensFromString(*reason)
				return sender(reason, nil)
			}
			// Reasoning ended, close the think block if open
			if pendingThinkClose {
				pendingThinkClose = false
				thinkClose := "</think>"
				if e := sender(&thinkClose, nil); e != nil {
					return e
				}
			}
			if delta != nil && *delta != "" {
				if *delta == "[DONE]" {
					return nil
				}
				roundTokens += tokenizer.NumTokensFromString(*delta)
				answer += *delta
				if e := sender(delta, nil); e != nil {
					return e
				}
			}
			return nil
		})
		// Close any unclosed think block after stream completes
		if pendingThinkClose {
			pendingThinkClose = false
			thinkClose := "</think>"
			if e := sender(&thinkClose, nil); e != nil {
				return totalTokens, e
			}
		}
		// Commit this round's token count to the running aggregate.
		// Prefer authoritative API usage over text-based estimation.
		commitRound(&cfg, roundTokens)
		if err != nil {
			return totalTokens, fmt.Errorf("round %d: %w", round, err)
		}

		var toolCalls []map[string]interface{}
		if cfg.ToolCallsResult != nil {
			toolCalls = *cfg.ToolCallsResult
		}

		if answer != "" && len(toolCalls) == 0 {
			return totalTokens, nil
		}
		if len(toolCalls) == 0 {
			return totalTokens, fmt.Errorf("round %d: no content and no tool_calls", round)
		}

		// A terminal tool's successful result is already the final answer:
		// stream it to the caller and stop instead of asking the model again
		// (Python chat_model.py:2574-2582). sendTerminal streams the result.
		// Streaming passes emptyTerminalIsHit=true: a Go streaming tool returns
		// "" to say "already delivered through the sink", and the loop must
		// stop even when the folded content is empty — unlike Python, whose rag
		// tool returns the full text and whose all-empty fallthrough would only
		// buy an extra model round whose output the mux drops.
		var termAnswer string
		var termHit bool
		history, termAnswer, termHit = appendToolResults(history, toolCalls, cm.ToolConfig.ToolCallSession, terminalSet(cm), true)
		if termHit {
			if err := sendTerminal(sender, &termAnswer); err != nil {
				return totalTokens, err
			}
			return totalTokens, nil
		}
		// history now carries this round's tool results; continue to the next round.
	}

	// Exceeded max rounds
	history = append(history, Message{
		Role:    "user",
		Content: fmt.Sprintf("Exceed max rounds: %d", maxRounds),
	})
	cfg := *chatCfg
	cfg.Stream = boolPtr(true)
	var exceedUsage TokenUsage
	cfg.UsageResult = &exceedUsage
	var exceedTokens int
	err := cm.ModelDriver.ChatStreamlyWithSender(ctx, *cm.ModelName, history, cm.APIConfig, &cfg, nil, func(delta *string, reason *string) error {
		if delta != nil && *delta != "" && *delta != "[DONE]" {
			exceedTokens += tokenizer.NumTokensFromString(*delta)
		}
		return nil
	})
	commitRound(&cfg, exceedTokens)
	return totalTokens, err
}

// appendToolResults executes tool calls concurrently, appends the assistant
// message with tool_calls and individual tool result messages to history.
//
// When terminal is non-empty, a successful call to one of those tools
// short-circuits: the loop returns (history, that result, true) so the caller
// treats it as the final answer instead of re-invoking the model (mirrors
// Python chat_model.py:619-627 / :2574-2582). An EMPTY terminal result does
// not win on its own: Python's fold only returns on a non-empty string
// (`if out:`, chat_model.py:696), so an empty result is skipped in favour of
// a later non-empty sibling. When NO terminal result is non-empty,
// emptyTerminalIsHit decides:
//   - false (non-streaming, Python :692-704 fallthrough): no hit — the tool
//     responses are in history and the loop runs another model round, which
//     then answers from them or re-calls the tool;
//   - true (streaming): hit with the first successful call's (empty) content.
//     The Go streaming tool returns "" to say "already delivered through the
//     sink", and the loop must stop (sendTerminal stays silent) instead of
//     paying another model round whose output the mux would only drop.
//
// When terminal is empty or no terminal tool fired, it returns
// (history, "", false) either way.
func appendToolResults(history []Message, toolCalls []map[string]interface{}, session ToolCallSession, terminal map[string]struct{}, emptyTerminalIsHit bool) ([]Message, string, bool) {
	if session == nil {
		history = append(history, Message{
			Role:      "assistant",
			Content:   nil,
			ToolCalls: toolCalls,
		})
		for _, tc := range toolCalls {
			tcID, _ := tc["id"].(string)
			history = append(history, Message{
				Role:       "tool",
				Content:    "Error: no tool session configured",
				ToolCallID: tcID,
			})
		}
		return history, "", false
	}
	var mu sync.Mutex
	var wg sync.WaitGroup
	type toolResult struct {
		index   int
		tcID    string
		name    string
		err     error
		content string
	}
	results := make([]toolResult, len(toolCalls))

	for i, tc := range toolCalls {
		wg.Add(1)
		go func(idx int, tcMap map[string]interface{}) {
			defer wg.Done()
			var result toolResult
			result.index = idx
			fn, ok := tcMap["function"].(map[string]interface{})
			if !ok {
				mu.Lock()
				results[idx] = result
				mu.Unlock()
				return
			}
			name, _ := fn["name"].(string)
			argsStr, _ := fn["arguments"].(string)
			result.tcID, _ = tcMap["id"].(string)
			result.name = name

			var args map[string]interface{}
			if err := json.Unmarshal([]byte(argsStr), &args); err != nil {
				args = map[string]interface{}{"raw_arguments": argsStr}
			}

			res, err := session.ToolCall(name, args)
			if err != nil {
				result.err = err
				result.content = fmt.Sprintf("Error: %s", err.Error())
			} else {
				result.content = res
			}
			mu.Lock()
			results[idx] = result
			mu.Unlock()
		}(i, tc)
	}
	wg.Wait()

	history = append(history, Message{
		Role:      "assistant",
		Content:   nil,
		ToolCalls: toolCalls,
	})

	// Every tool_call the assistant declared gets its matching tool message,
	// well-formed history whichever path the caller takes. Then two passes over
	// the (tiny) results slice pick the fold: the FIRST non-empty successful
	// terminal result wins (Python chat_model.py:696 `if out:` — an empty
	// terminal result never ships while a non-empty sibling exists, trimmed
	// whitespace counts as empty), then the empty fallthrough per
	// emptyTerminalIsHit (see the doc comment above).
	for _, r := range results {
		history = append(history, Message{
			Role:       "tool",
			Content:    r.content,
			ToolCallID: r.tcID,
		})
	}
	var terminalAnswer string
	var terminalHit bool
	for _, r := range results {
		if r.err == nil && r.name != "" && strings.TrimSpace(r.content) != "" {
			if _, ok := terminal[r.name]; ok {
				terminalAnswer, terminalHit = r.content, true
				break
			}
		}
	}
	if !terminalHit && emptyTerminalIsHit {
		for _, r := range results {
			if r.err == nil && r.name != "" {
				if _, ok := terminal[r.name]; ok {
					terminalAnswer, terminalHit = r.content, true
					break
				}
			}
		}
	}

	// A terminal tool short-circuits on success regardless of what it returned.
	// Tools that stream their own answer through the caller's sender return an
	// empty string to say "nothing more to emit", and sendTerminal's empty guard
	// then keeps the loop from re-streaming it (mirrors Python, where the
	// terminal result is ignored once the inner answer_sink has streamed).
	if terminalHit {
		return history, terminalAnswer, true
	}
	return history, "", false
}

func boolPtr(b bool) *bool {
	return &b
}

// sendTerminal streams a terminal tool's already-final answer to the sender as
// plain text deltas (no reasoning), mirroring how Python yields the terminal
// result and stops.
func sendTerminal(sender func(*string, *string) error, answer *string) error {
	if answer == nil || *answer == "" {
		return nil
	}
	return sender(answer, nil)
}