ton/YiemAgent
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 6 Wiki Activity

update

Browse Source
This commit is contained in:
narawat
2026-07-20 12:23:03 +07:00
parent ddbb135b6b
commit ef523aaa48
4 changed files with 284 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llmfunction.jl
+237 -8
View File
@@ -5,7 +5,7 @@ export virtualWineUserChatbox, jsoncorrection, search_wine_database!, # recomme
extractWineAttributes_2, paraphrase, SQLexecution
using HTTP, JSON, URIs, Random, PrettyPrinting, UUIDs, Dates, DataFrames, DataStructures,
Base64
Base64, Serde
using GeneralUtils, SQLLLM
using ..type, ..util
@@ -277,14 +277,19 @@ end
# Example
```jldoctest
julia> using ChatAgent
julia> agent = ChatAgent.agentReflex("Jene")
julia> input = "{\"food\": \"pizza\", \"occasion\": \"anniversary\"}"
julia> result = checkinventory(agent, input)
"{"wine 1": {\"Winery\": \"Pichon Baron\", \"wine name\": \"Pauillac (Grand Cru Classé)\", \"grape variety\": \"Cabernet Sauvignon\", \"year\": 2010, \"price\": \"125 USD\", \"stock ID\": \"ar-17\"}, }"
julia> agent = YiemAgent.sommelier(...)
julia> thoughtdict =
OrderedDict{String, Any}(
"plan" => "The user is asking a very specific question about a wine (Brunello di Montalcino from Tenuta CastelGiocondo). Although the policy suggests gathering budget, wine type, and occasion, the user has provided enough specific information (name, region, producer) to attempt a direct search in the database. I will use the SEARCH_WINE_DATABASE action to check if this specific wine is in our inventory.",
"action_name" => "SEARCH_WINE_DATABASE",
"action_input" => "Brunello di Montalcino from Tenuta CastelGiocondo")
```
"""
function search_wine_database!(a::T, thoughtdict::AbstractDict; useSQLLLM::Bool=false
)::NamedTuple{(:thoughtdict, :result_raw), Tuple{OrderedDict, Any}} where {T<:agent}
# WORKING
# look_for_wine_in_wine_database(a, thoughtdict["action_input"])
println("\ncheckinventory order: $(thoughtdict["action_input"]) ", @__FILE__, ":", @__LINE__, " $(Dates.now())")
wineattributes_1 = extractWineAttributes_1(a, thoughtdict["action_input"])
@@ -366,9 +371,9 @@ function generatesql(a::T, searchterm::String,
Fulfill the objective.
# you should then respond to the user with interleaving plan, action_name, action_input
1) "plan, Based on the current situation, state a complete action plan to complete the task and rationale. Be specific.
2) "action_name, Must be "RUNSQL"
3) "action_input, The input to the action you are about to perform according to your plan.
1) "plan", Based on the current situation, state a complete action plan to complete the task and rationale. Be specific.
2) "action_name", Must be "RUNSQL"
3) "action_input", The input to the action you are about to perform according to your plan.
After the action is executed you gets "action_result". It is the output from the action you selected.
# you should only respond in JSON format as described below
@@ -610,6 +615,230 @@ function generatesql(a::T, searchterm::String,
error("SQLLLM DecisionMaker() failed to generate a thought \n", response)
end
"""
# Example
```jldoctest
julia> using ChatAgent
julia> agent = YiemAgent.sommelier(...)
julia> thoughtdict =
OrderedDict{String, Any}(
"plan" => "The user is asking a very specific question about a wine (Brunello di Montalcino from Tenuta CastelGiocondo). Although the policy suggests gathering budget, wine type, and occasion, the user has provided enough specific information (name, region, producer) to attempt a direct search in the database. I will use the SEARCH_WINE_DATABASE action to check if this specific wine is in our inventory.",
"action_name" => "SEARCH_WINE_DATABASE",
"action_input" => "Brunello di Montalcino from Tenuta CastelGiocondo")
```
julia> look_for_wine_in_wine_database(agent, thoughtdict["action_input"])
"""
function look_for_wine_in_wine_database(a::T, searchterm::String,
; maxattempt=10
)::String where {T<:agent}
systemmsg =
"""
# situation
At each round of conversation, you will be given the following:
- user search term
- database tables schema
# objective
Consult the provided database schema (tables and columns), please map a user's natural-language search term to the appropriate database columns and tables—identify the relevant fields, operators, and values (e.g., for SQL filtering).
# your responsibility includes
Fulfill the objective.
# you should only respond in YAML format as described below
table_name_1:
column_name_1:
operator: "="
value: "..."
column_name_2:
operator: "="
value: "..."
...
table_name_2:
column_name_1:
operator: "="
value: "..."
column_name_2:
operator: "="
value: "..."
...
# here are some example
<user>
4-wheel drive car with red color that will give me fast and furious emotion. No more than 7000 USD
</user>
<assistant>
car_info: # table_name
drive_type: # column_name
operator: "=" # operator is not "N/A" because drive_type column store quantitative value
value: "4-wheel" # column_value
color:
operator: "=" # operator is not "N/A" because color column store quantitative value
value: "red"
drive_feeling:
operator: "N/A" # operator is "N/A" because drive_feeling column store qualitative value
value: "fast and furious"
price_list:
price:
operator: "<" # operator is not "N/A" because drive_type column store quantitative value
value: "7000"
</assistant>
"""
# use find_related_tables_for_user_question and inject only related table schema for a given search term
# to LLM instead of giving LLM all tables schema.
related_tables = a.context.find_related_tables_for_user_question(searchterm)
table_schema = ""
for table in related_tables
_table_schema_str = GeneralUtils.get_db_table_schema_simple(a.context.pg_conn_str, table)
table_schema_str = sprint(show, _table_schema_str) * "\n"
table_schema = table_schema * table_schema_str
end
context =
"""
<internal_context_for_assistant>
<database_table_schema>
$table_schema
</database_table_schema>
</internal_context_for_assistant>
"""
input = context * searchterm
msg = Dict(
"model" => "gemma-4-E4B-it-UD-Q4_K_XL",
"messages" => [
Dict(
"role" => "system",
"content" => [
Dict("type" => "text", "text" => systemmsg),
]
),
Dict(
"role" => "user",
"content" => [
Dict("type" => "text", "text" => input),
]
),
],
"temperature" => 0.7
)
for attempt in 1:maxattempt
response = a.context.text2textInstructLLM("random_id", msg)
responsedict = Serde.parse_yaml(response)
"""
responsedict = Dict(
"wine" => Dict(
"tasting_notes" => Dict(
"operator" => "N/A",
"value" => "casual dinner"
),
"wine_type" => Dict(
"operator" => "=",
"value" => "red"
)
),
"retailer_wine" => Dict(
"currency" => Dict(
"operator" => "=", "value" => "USD"
),
"price" => Dict(
"operator" => "<", "value" => "1000"
)
)
)
"""
for (table_name, table_info_dict) in responsedict
for (column_name, v) in table_info_dict
#
do_not_resolve_BM25_list = ["tasting_notes", "seo_name", "vintage", "grape"]
if column_name ∉ do_not_resolve_list
words_catalog = GeneralUtils.harvest_entity_catalog(a.context.pg_conn_str, table_name, column_name)
resolved_word = GeneralUtils.resolve_entity(v["value"], words_catalog; threshold=0.9)
table_info_dict[column_name] = resolved_word
end
end
end
# check each attributes against each column in a database table with BM25 and get the closest
# word match because there is a typo sometimes.
for (k, v) in responsedict
if k ∉ ["tasting_notes"]
words_catalog = GeneralUtils.harvest_entity_catalog(a.context.pg_conn_str, "wine", k)
resolved_word = GeneralUtils.resolve_entity(v, words_catalog; threshold=0.9)
responsedict[k] = resolved_word
end
end
#WORKING
println("\n", responsedict)
@info "test done " @__LINE__
error(9999)
think, response = GeneralUtils.extractthink(response)
responsedict = nothing
try
_responsedict = JSON.parse(response)
responsedict = GeneralUtils.dictify(_responsedict, keytype=String)
catch
println("\nERROR decisionMaker() failed to parse response: $response ", @__FILE__, ":", @__LINE__, " $(Dates.now())")
continue
end
# check each attributes against each column in a database table with BM25 and get the closest
# word match because there is a typo sometimes.
for (k, v) in responsedict
if k ∉ ["tasting_notes"]
words_catalog = GeneralUtils.harvest_entity_catalog(a.context.pg_conn_str, "wine", k)
resolved_word = GeneralUtils.resolve_entity(v, words_catalog; threshold=0.9)
responsedict[k] = resolved_word
end
end
# LLM already extract user search term against tables schema
# Ex. responsedict = Dict(
# "wine_type"=> "red", # hard constraint
# "region"=> "bordeaux", # hard constraint
# "price_max"=> "100", # hard constraint
# "tasting_notes"=> "fruity, oak" # semantic search)
return items
end
error("SQLLLM DecisionMaker() failed to generate a thought \n", response)
end
function SQLexecution(executeSQL::Function, sql::T
)::NamedTuple where {T<:AbstractString}