9iQ X d Z ddlZddlZddlmZmZmZ ddlmZ ddl m
Z
m Z m Z ddl
mZmZmZmZ ddlmZ ddlmZmZ dd lmZ dd
lmZ dd lm Z m Z dd l m Z m Z eee ge!f Z"d
Z#dZ$dZ%dZ&e'e d e(f Z) e'e d e!f Z* e'e d e!f Z+ e)e*z e+z Z, de de"fdZ- G d de Z.y)zSummarization middleware. N)CallableIterableMapping)partial)AnyLiteralcast)
AnyMessageMessageLikeRepresentation
RemoveMessage ToolMessage HumanMessage)count_tokens_approximately
trim_messages)REMOVE_ALL_MESSAGES)Runtime)AgentMiddleware
AgentState)
BaseChatModelinit_chat_modela
Context Extraction Assistant
Your sole objective in this task is to extract the highest quality/most relevant context from the conversation history below.
You're nearing the total number of input tokens you can accept, so you must extract the highest quality/most relevant pieces of information from your conversation history.
This context will then overwrite the conversation history presented below. Because of this, ensure the context you extract is only the most important information to your overall goal.
The conversation history below will be replaced with the context you extract in this step. Because of this, you must do your very best to extract and record all of the most important context from the conversation history.
You want to ensure that you don't repeat any actions you've already completed, so the context you extract from the conversation history should be focused on the most important information to your overall goal.
The user will message you with the full message history you'll be extracting context from, to then replace. Carefully read over it all, and think deeply about what information is most important to your overall goal that should be saved:
With all of this in mind, please carefully read over the entire conversation history, and extract the most important and relevant context to replace it so that you can free up space in the conversation history.
Respond ONLY with the extracted context. Do not include any additional information, or text before or after the extracted context.
Messages to summarize:
{messages}
i fractiontokensmessagesmodelreturnc N | j dk( rt t d S t S )zATune parameters of approximate token counter based on model type.zanthropic-chatgffffff
@)chars_per_token) _llm_typer r )r s s/var/www/html/backtest/airagagent/rag_env/lib/python3.12/site-packages/langchain/agents/middleware/summarization.py _get_approximate_token_counterr# y s% ** 13GG %% c e Zd ZdZddefeeeddee z de
e e
z dz de
de d ed
e
dz d ed df fd
Zdeded eeef dz fdZdeded eeef dz fdZde e de
d efdZde e d e
fdZde e d e
dz fdZd e
dz fdZde
ded e
fdZded e e fdZ de e d dfd Z d e e d e
d e e e e e f fd Z de e d e
d e
fd!Z!de e d e
d e
fd"Z"d#e e d efd$Z#d#e e d efd%Z$de e d e e fd&Z% xZ&S )'SummarizationMiddlewarea; Summarizes conversation history when token limits are approached.
This middleware monitors message token counts and automatically summarizes older
messages when a threshold is reached, preserving recent messages and maintaining
context continuity by ensuring AI/Tool message pairs remain together.
Nr )triggerkeep
token_countersummary_prompttrim_tokens_to_summarizer r' r( r) r* r+ deprecated_kwargsr c @ d|v r)|d }t j dt d ||d|f}d|v r0|d }t j dt d |d t fk( rd |f}t | t
|t r t | }|| _ |
d| _
g } nUt
|t r)|D
cg c] }
| j |
d
} }
| | _
| } n | j |d
} | | _
| g} | | _
| j |d | _ |t u rt! | j | _ n|| _ || _ || _ t) d | j D }
| j d
dk( rd}
|
r | j+
d}t- | yyc c}
w )a Initialize summarization middleware.
Args:
model: The language model to use for generating summaries.
trigger: One or more thresholds that trigger summarization.
Provide a single
[`ContextSize`][langchain.agents.middleware.summarization.ContextSize]
tuple or a list of tuples, in which case summarization runs when any
threshold is met.
!!! example
```python
# Trigger summarization when 50 messages is reached
("messages", 50)
# Trigger summarization when 3000 tokens is reached
("tokens", 3000)
# Trigger summarization either when 80% of model's max input tokens
# is reached or when 100 messages is reached (whichever comes first)
[("fraction", 0.8), ("messages", 100)]
```
See [`ContextSize`][langchain.agents.middleware.summarization.ContextSize]
for more details.
keep: Context retention policy applied after summarization.
Provide a [`ContextSize`][langchain.agents.middleware.summarization.ContextSize]
tuple to specify how much history to preserve.
Defaults to keeping the most recent `20` messages.
Does not support multiple values like `trigger`.
!!! example
```python
# Keep the most recent 20 messages
("messages", 20)
# Keep the most recent 3000 tokens
("tokens", 3000)
# Keep the most recent 30% of the model's max input tokens
("fraction", 0.3)
```
token_counter: Function to count tokens in messages.
summary_prompt: Prompt template for generating summaries.
trim_tokens_to_summarize: Maximum tokens to keep when preparing messages for
the summarization call.
Pass `None` to skip trimming entirely.
max_tokens_before_summaryzOmax_tokens_before_summary is deprecated. Use trigger=('tokens', value) instead. )
stacklevelNr messages_to_keepzEmessages_to_keep is deprecated. Use keep=('messages', value) instead.r r' r( c 3 , K | ] }|d dk( yw)r r N ).0 conditions r" z3SummarizationMiddleware.__init__.. s diy|z 9 ds r r Ta Model profile information is required to use fractional token limits, and is unavailable for the specified model. Please use absolute token counts instead, or pass `
ChatModel(..., profile={"max_input_tokens": ...})`.
with a desired integer value of the model's maximum input tokens.)warningswarnDeprecationWarning_DEFAULT_MESSAGES_TO_KEEPsuper__init__
isinstancestrr r r' list_validate_context_size_trigger_conditionsr( r r# r) r* r+ any_get_profile_limits
ValueError)selfr r' r( r) r* r+ r, valuetrigger_conditionsitemvalidated_list validatedrequires_profilemsg __class__s r" r< z SummarizationMiddleware.__init__ s F '*; ;%&ABE MMa"
5#4#U+ !2 2%&89E MMW"
$=>>"E*
eS !#E*E
?CGDL46
&W^ _td 9 9$ J _N _)DL!/ 33GYGI$DL"+ #5 //f= 6 6!?
!KD !.D ,(@% d4KcKc dd 99Q<: % # 8 8 : BT
S/ ! !C ) `s 3Fstateruntimec @ |d }| j | | j | }| j || sy| j | }|dk ry| j || \ }}| j | }| j
| } dt t g| |iS zOProcess messages before model invocation, potentially triggering summarization.r Nr )id) _ensure_message_idsr) _should_summarize_determine_cutoff_index_partition_messages_create_summary_build_new_messagesr r
rE rN rO r total_tokens cutoff_indexmessages_to_summarizepreserved_messagessummary new_messagess
r" before_modelz$SummarizationMiddleware.before_model s $ *))(3 %%h =33H= 1 484L4LXWc4d1 1&&'<=//8
!45 $
r$ c \ K |d }| j | | j | }| j || sy| j | }|dk ry| j || \ }}| j | d{ }| j
| } dt t g| |iS 7 ,wrQ ) rS r) rT rU rV _acreate_summaryrX r r rY s
r"
abefore_modelz%SummarizationMiddleware.abefore_model$ s $ *))(3 %%h =33H= 1 484L4LXWc4d1 1 --.CDD//8
!45 $
Es A;B,=B*>-B,rZ c | j sy| j D ][ \ }}|dk( rt | |k\ r y|dk( r||k\ r y|dk( s-| j }|@t ||z }|dk rd}||k\ s[ y y)zGDetermine whether summarization should run for the current token usage.Fr Tr r r )rA lenrC int)rE r rZ kindrF max_input_tokens thresholds r" rT z)SummarizationMiddleware._should_summarize? s '' 33
KD%z!c(mu&<x LE$9z!#'#;#;#= #+ 05 8 9 > !I 9,
r$ c | j \ }}|dv r+| j | }||S | j |t S | j |t d| S )z7Choose cutoff index respecting retention configuration.> r r rg )r( _find_token_based_cutoff_find_safe_cutoffr: r )rE r rh rF token_based_cutoffs r" rU z/SummarizationMiddleware._determine_cutoff_indexT se ii e ) )!%!>!>x!H !-)) ))(4MN N%%hUE0BCCr$ c R |sy| j \ }}|dk( r"| j }|yt ||z }n|dk( r t | }ny|dk rd}| j | |k rydt | }}t | }t | j dz } t
| D ]2 }
||k\ r n+||z dz } | j || d |k r| }| }.| dz }4 |t | k( r|}|t | k\ r t | dk( ryt | dz
}| j || S )z2Find cutoff index based on target token retention.r r Nr re r/ )r( rC rg r) rf
bit_lengthrange_find_safe_cutoff_point) rE r rh rF ri target_token_countleftrightcutoff_candidatemax_iterations_mids r" rl z0SummarizationMiddleware._find_token_based_cutoff` s^ ii e : # 7 7 9 '!$%5%=!>
X
!$U "!" h '+= = X e x= X 113a7~& Au}% r r z( thresholds must be greater than 0, got z Unsupported context size type z for )rD )rE r~ r rh rF rL s r" r@ z.SummarizationMiddleware._validate_context_size s e : u>>#N#33XY^X__`a o% +
+z'((PQVPWWXY o% 34&n=MQOC S/ !r$ r^ c " t d| gS )Nz0Here is a summary of the conversation to date:
)contentr )rE r^ s r" rX z+SummarizationMiddleware._build_new_messages s #UV]U^!_ `
r$ c p |D ]1 }|j t t j |_ 3 y)zAEnsure all messages have unique IDs for the add_messages reducer.N)rR r> uuiduuid4)rE r rL s r" rS z+SummarizationMiddleware._ensure_message_ids s, +Cvv~ TZZ\* +r$ conversation_messagesr[ c |d| }||d }||fS )zAPartition messages into those to summarize and those to preserve.Nr3 )rE r r[ r\ r] s r" rV z+SummarizationMiddleware._partition_messages s* !6m| D 2<= A$&888r$ r1 c ` t | |k ryt | |z
}| j || S )a Find safe cutoff point that preserves AI/Tool message pairs.
Returns the index where messages can be safely cut without separating
related AI and Tool messages. Returns `0` if no safe cutoff is found.
This is aggressive with summarization - if the target cutoff lands in the
middle of tool messages, we advance past all of them (summarizing more).
r )rf rr )rE r r1
target_cutoffs r" rm z)SummarizationMiddleware._find_safe_cutoff s6 x= , ,H
(88
++HmDDr$ c |t | k r:t || t r'|dz
}|t | k rt || t r'|S )a Find a safe cutoff point that doesn't split AI/Tool message pairs.
If the message at cutoff_index is a ToolMessage, advance until we find
a non-ToolMessage. This ensures we never cut in the middle of parallel
tool call responses.
re )rf r= r
)rE r r[ s r" rr z/SummarizationMiddleware._find_safe_cutoff_point sL S]*z(<:PR]/^ A L S]*z(<:PR]/^r$ r\ c |sy| j | }|sy | j j | j j | }|j
j
S # t $ r}d|cY d}~S d}~ww xY w)(Generate summary for the given messages.!No previous conversation history.0Previous conversation was too long to summarize.r Error generating summary: N)_trim_messages_for_summaryr invoker* formattextstrip ExceptionrE r\ trimmed_messagesresponsees r" rW z'SummarizationMiddleware._create_summary s~ $6 ::;PQ E 6zz(()<)<)C)CM])C)^_H==&&( ( 6/u5 5 6s AA( ( B 1A;5B ;B c " K |sy| j | }|sy | j j | j j | d{ }|j
j
S 7 # t $ r}d|cY d}~S d}~ww xY ww)r r r r Nr )r r ainvoker* r r r r r s r" rb z(SummarizationMiddleware._acreate_summary s $6 ::;PQ E 6 !ZZ / /##**4D*E H ==&&( ( 6/u5 5 6sF B8A4 A2 A4 1B2A4 4 B =BB BB Bc | j |S t dt || j | j dddd S # t $ r
|t
d cY S w xY w)z6Trim messages to fit within summary generation limits.Nzlist[AnyMessage]humanlastT)
max_tokensr) start_onstrategy
allow_partialinclude_system)r+ r r r) r _DEFAULT_FALLBACK_MESSAGE_COUNT)rE r s r" r z2SummarizationMiddleware._trim_messages_for_summary st ?,,4 " # < <"&"4"4 $ #"&#'
? < < => > ?s
A /A AA)'__name__
__module__ __qualname____doc__r: r DEFAULT_SUMMARY_PROMPT_DEFAULT_TRIM_TOKEN_LIMITr> r ContextSizer? TokenCounterrg r r< r r dictr` rc r
boolrT rU rl rC r@ r rX rS tuplerV rm rr rW rb r
__classcell__)rM s @r" r& r& sh ;? ')B C&@ 4/H}"]"}" tK 0047 }"
}" $
}" }" #&*}" !}"
}"~
*
w
4S>TXCX
6
g
$sTWx.[_J_
6$z*: # RV *
DZ0@
DS
D/Hj1A /HcDj /Hb S4Z "k 3 S^
3
4 3E
+D,< + + 9 #J / 9 9
tJ j!11 2 9
E$z*:
Ec
EVY
E Z0@ PS X[
6T*5E
6#
6 6Dr s 7 7 % % 7 S & I @";<=sBC 8 "$
+U23 gh',-
+S01
-? 8&- &L &R?o R?r$