mirror of
https://github.com/infiniflow/ragflow.git
synced 2026-07-03 01:01:56 +08:00
742188c3bb
### What problem does this PR solve?
_Briefly describe what this PR aims to solve. Include background context
that will help reviewers understand the purpose of the PR._
### Type of change
- [x] Bug Fix (non-breaking change which fixes an issue)
- [x] New Feature (non-breaking change which adds functionality)
- [x] Other (please describe):
## Summary
Agent (Canvas) runs previously did not surface token usage in the SSE
stream, and RAGFlow's own Langfuse generations for agent runs were
missing the prompt/completion split and the session/user correlation.
This made it impossible for an external caller (or Langfuse) to
reconcile an agent turn's cost with the upstream provider (e.g.
OpenRouter), because a single turn can issue several distinct LLM calls
(query rewriting / cross-language translation, multi-round tool
reasoning, nested sub-agents, and the final answer).
This PR introduces a per-run token usage sink so that **every** LLM call
in a run is aggregated and reported once, and enriches Langfuse
generations with the prompt/completion split plus session/user
attributes.
## What changes
### 1. Per-run token usage sink (`common/token_utils.py`)
- Adds two `contextvars`: `token_usage_sink` (a mutable per-run
accumulator) and `langfuse_run_attrs` (session_id/user_id for the run).
- Adds `record_run_token_usage(...)` (thread-safe via a lock, because
`thread_pool_exec` copies the context into worker threads that share the
sink dict) and `usage_from_response(...)` which extracts a
`{prompt_tokens, completion_tokens, total_tokens}` split from
OpenAI/OpenRouter-style responses.
### 2. Provider layer captures the prompt/completion split
(`rag/llm/chat_model.py`)
- `LiteLLMBase` and `Base` now store `self.last_usage`
(prompt/completion/total) for the most recent chat call, in both the
plain and tool-calling paths.
- Streaming requests set `stream_options.include_usage = True` (LiteLLM
path) so the authoritative usage arrives on the final chunk; this is
read even on the usage-only chunk that carries no `choices`.
- Fixes a multi-round accounting bug in `*_with_tools`: token totals
were **overwritten** by each round (`total_tokens = tol`) instead of
accumulated, undercounting multi-round tool conversations. Each round is
now committed to a running aggregate.
### 3. LLMBundle reports usage once, per call
(`api/db/services/llm_service.py`)
- New `_report_usage(total_tokens)` records the call's usage into the
active run sink and returns the prompt/completion/total split for
Langfuse. The split is only used when it is consistent with the
authoritative total; otherwise only the total is reported.
- All three chat entry points (`async_chat`, `async_chat_streamly`,
`async_chat_streamly_delta`) now emit `usage_details` with
`input`/`output`/`total` instead of total-only.
- `_start_langfuse_observation` now applies `session_id`/`user_id` from
the per-run context (`langfuse_run_attrs`) so agent-run generations are
correctly grouped, even though agent LLMBundles are constructed without
those attributes.
### 4. Canvas installs the sink and emits the aggregate
(`agent/canvas.py`)
- `Canvas.run()` installs a fresh `token_usage_sink` and
`langfuse_run_attrs` (from `user_id`/`session_id`) at the start of every
turn.
- `message_end` now includes an aggregated `usage` object:
`{prompt_tokens, completion_tokens, total_tokens, calls}` covering all
LLM calls in the run.
### 5. Pass session id into the run
(`api/db/services/canvas_service.py`)
- `completion()` forwards `session_id` to `Canvas.run()` for Langfuse
session correlation.
## Why a context variable
LLM calls in an agent run originate from many places that each build
their own `LLMBundle` (e.g. `cross_languages`/`keyword_extraction`
helpers, the Agent component, and nested sub-agents invoked as tools). A
run-scoped context variable is the only non-invasive chokepoint that
captures all of them exactly once, including nested agents (which run in
the same async context) and thread-pool tools (the executor copies the
context).
## Behavior / compatibility
- No public API or wire-format removal: `message_end` gains an
additional optional `usage` field; existing consumers are unaffected.
- When a provider does not return authoritative usage, behavior falls
back to the previous token estimate (total only, no split).
- Non-agent flows (Dataflow `Pipeline`, sync `Graph.run`) are untouched.
## Testing
- [x] Simple agent answer: `message_end.usage.total_tokens` matches
provider usage.
- [x] Agent with cross-language retrieval: aggregate equals the sum of
both provider calls.
- [x] Tool-calling agent (multi-round): total accumulates across rounds.
- [x] Nested agent (agent-as-tool): sub-agent tokens included in the
parent run total.
- [x] Langfuse: agent generations show input/output split and are
grouped by session/user.
---------
Co-authored-by: yzc <yuzhichang@gmail.com>
Co-authored-by: Cursor <cursoragent@cursor.com>