From 484072b448b5a60fa6398028882b3424b09b20d1 Mon Sep 17 00:00:00 2001 From: Andrew Macri Date: Fri, 22 Sep 2023 16:51:33 -0600 Subject: [PATCH] ## [Security Solution] [Elastic AI Assistant] LangChain Agents and Tools integration for ES|QL query generation via ELSER This PR integrates [LangChain](https://www.langchain.com/) [Agents](https://js.langchain.com/docs/modules/agents/) and [Tools](https://js.langchain.com/docs/modules/agents/tools/) with the [Elastic AI Assistant](https://www.elastic.co/blog/introducing-elastic-ai-assistant). These abstractions enable the LLM to dynamically choose whether or not to query, via [ELSER](https://www.elastic.co/guide/en/machine-learning/current/ml-nlp-elser.html), an [ES|QL](https://www.elastic.co/blog/elasticsearch-query-language-esql) knowledge base. Context from the knowledge base is used to generate `ES|QL` queries, or answer questions about `ES|QL`. Registration of the tool occurs in `x-pack/plugins/elastic_assistant/server/lib/langchain/execute_custom_llm_chain/index.ts`: ```typescript const tools: Tool[] = [ new ChainTool({ name: 'esql-language-knowledge-base', description: 'Call this for knowledge on how to build an ESQL query, or answer questions about the ES|QL query language.', chain, }), ]; ``` The `tools` array above may be updated in future PRs to include, for example, an `ES|QL` query validator endpoint. ### Details The `callAgentExecutor` function in `x-pack/plugins/elastic_assistant/server/lib/langchain/execute_custom_llm_chain/index.ts`: 1. Creates a `RetrievalQAChain` from an `ELSER` backed `ElasticsearchStore`, which serves as a knowledge base for `ES|QL`: ```typescript // ELSER backed ElasticsearchStore for Knowledge Base const esStore = new ElasticsearchStore(esClient, KNOWLEDGE_BASE_INDEX_PATTERN, logger); const chain = RetrievalQAChain.fromLLM(llm, esStore.asRetriever()); ``` 2. Registers the chain as a tool, which may be invoked by the LLM based on its description: ```typescript const tools: Tool[] = [ new ChainTool({ name: 'esql-language-knowledge-base', description: 'Call this for knowledge on how to build an ESQL query, or answer questions about the ES|QL query language.', chain, }), ]; ``` 3. Creates an Agent executor that combines the `tools` above, the `ActionsClientLlm` (an abstraction that calls `actionsClient.execute`), and memory of the previous messages in the conversation: ```typescript const executor = await initializeAgentExecutorWithOptions(tools, llm, { agentType: 'chat-conversational-react-description', memory, verbose: false, }); ``` Note: Set `verbose` above to `true` to for detailed debugging output from LangChain. 4. Calls the `executor`, kicking it off with `latestMessage`: ```typescript await executor.call({ input: latestMessage[0].content }); ``` ### Changes to `x-pack/packages/kbn-elastic-assistant` A client side change was required to the assistant, because the response returned from the agent executor is JSON. This response is parsed on the client in `x-pack/packages/kbn-elastic-assistant/impl/assistant/api.tsx`: ```typescript return assistantLangChain ? getFormattedMessageContent(result) : result; ``` Client-side parsing of the response only happens when then `assistantLangChain` feature flag is `true`. ## Desk testing Set ```typescript assistantLangChain={true} ``` in `x-pack/plugins/security_solution/public/assistant/provider.tsx` to enable this experimental feature in development environments. Also (optionally) set `verbose` to `true` in the following code in ``x-pack/plugins/elastic_assistant/server/lib/langchain/execute_custom_llm_chain/index.ts``: ```typescript const executor = await initializeAgentExecutorWithOptions(tools, llm, { agentType: 'chat-conversational-react-description', memory, verbose: true, }); ``` After setting the feature flag and optionally enabling verbose debugging output, you may ask the assistant to generate an `ES|QL` query, per the example in the next section. ### Example output When the Elastic AI Assistant is asked: ``` From employees, I want to see the 5 earliest employees (hire_date), I want to display only the month and the year that they were hired in and their employee number (emp_no). Format the date as e.g. "September 2019". Only show the query ``` it replies: ``` Here is the query to get the employee number and the formatted hire date for the 5 earliest employees by hire_date: FROM employees | KEEP emp_no, hire_date | EVAL month_year = DATE_FORMAT(hire_date, "MMMM YYYY") | SORT hire_date | LIMIT 5 ``` Per the screenshot below: ![ESQL_query_via_langchain_agents_and_tools](https://github.com/elastic/kibana/assets/4459398/c5cc75da-f7aa-4a12-9078-ed531f3463e7) The `verbose: true` output from LangChain logged to the console reveals that the prompt sent to the LLM includes text like the following: ``` Assistant can ask the user to use tools to look up information that may be helpful in answering the users original question. The tools the human can use are:\\n\\nesql-language-knowledge-base: Call this for knowledge on how to build an ESQL query, or answer questions about the ES|QL query language. ``` along with instructions for "calling" the tool like a function. The debugging output also reveals the agent selecting the tool, and returning results from ESLR: ``` [agent/action] [1:chain:AgentExecutor] Agent selected action: { "tool": "esql-language-knowledge-base", "toolInput": "Display the 'emp_no', month and year of the 5 earliest employees by 'hire_date'. Format the date as 'Month Year'.", "log": "```json\n{\n \"action\": \"esql-language-knowledge-base\",\n \"action_input\": \"Display the 'emp_no', month and year of the 5 earliest employees by 'hire_date'. Format the date as 'Month Year'.\"\n}\n```" } [tool/start] [1:chain:AgentExecutor > 4:tool:ChainTool] Entering Tool run with input: "Display the 'emp_no', month and year of the 5 earliest employees by 'hire_date'. Format the date as 'Month Year'." [chain/start] [1:chain:AgentExecutor > 4:tool:ChainTool > 5:chain:RetrievalQAChain] Entering Chain run with input: { "query": "Display the 'emp_no', month and year of the 5 earliest employees by 'hire_date'. Format the date as 'Month Year'." } [retriever/start] [1:chain:AgentExecutor > 4:tool:ChainTool > 5:chain:RetrievalQAChain > 6:retriever:VectorStoreRetriever] Entering Retriever run with input: { "query": "Display the 'emp_no', month and year of the 5 earliest employees by 'hire_date'. Format the date as 'Month Year'." } [retriever/end] [1:chain:AgentExecutor > 4:tool:ChainTool > 5:chain:RetrievalQAChain > 6:retriever:VectorStoreRetriever] [115ms] Exiting Retriever run with output: { "documents": [ { "pageContent": "[[esql-date_format]]\n=== `DATE_FORMAT`\nReturns a string representation of a date in the provided format. If no format\nis specified, the `yyyy-MM-dd'T'HH:mm:ss.SSSZ` format is used.\n\n[source,esql]\n----\nFROM employees\n| KEEP first_name, last_name, hire_date\n| EVAL hired = DATE_FORMAT(hire_date, \"YYYY-MM-dd\")\n----\n", ``` The documents containing `ES|QL` examples, retrieved from ELSER, are sent back to the LLM to answer the original question, per the abridged output below: ``` [llm/start] [1:chain:AgentExecutor > 4:tool:ChainTool > 5:chain:RetrievalQAChain > 7:chain:StuffDocumentsChain > 8:chain:LLMChain > 9:llm:ActionsClientLlm] Entering LLM run with input: { "prompts": [ "Use the following pieces of context to answer the question at the end. If you don't know the answer, just say that you don't know, don't try to make up an answer.\n\n[[esql-date_format]]\n=== `DATE_FORMAT`\nReturns a string representation of a date in the provided format. If no format\nis specified, the `yyyy-MM-dd'T'HH:mm:ss.SSSZ` format is used.\n\n[source,esql]\n----\nFROM employees\n| KEEP first_name, last_name, hire_date\n| EVAL hired = DATE_FORMAT(hire_date, \"YYYY-MM-dd\")\n----\n\n\n[[esql-date_trunc]]\n=== `DATE_TRUNC`\nRounds down a date to the closest interval. Intervals can be expressed using the\n<>.\n\n[source,esql]\n----\nFROM employees\n| EVAL year_hired = DATE_TRUNC(1 year, hire_date)\n| STATS count(emp_no) BY year_hired\n| SORT year_hired\n----\n\n\n[[esql-from]]\n=== `FROM`\n\nThe `FROM` source command returns a table with up to 10,000 documents from a\ndata stream, index, ``` ### Complete (verbose) LangChain output from the example The following `verbose: true` output from LangChain below was produced via the example in the previous section: ``` [chain/start] [1:chain:AgentExecutor] Entering Chain run with input: { "input": "\n\n\n\nFrom employees, I want to see the 5 earliest employees (hire_date), I want to display only the month and the year that they were hired in and their employee number (emp_no). Format the date as e.g. \"September 2019\". Only show the query", "chat_history": [] } [chain/start] [1:chain:AgentExecutor > 2:chain:LLMChain] Entering Chain run with input: { "input": "\n\n\n\nFrom employees, I want to see the 5 earliest employees (hire_date), I want to display only the month and the year that they were hired in and their employee number (emp_no). Format the date as e.g. \"September 2019\". Only show the query", "chat_history": [], "agent_scratchpad": [], "stop": [ "Observation:" ] } [llm/start] [1:chain:AgentExecutor > 2:chain:LLMChain > 3:llm:ActionsClientLlm] Entering LLM run with input: { "prompts": [ "[{\"lc\":1,\"type\":\"constructor\",\"id\":[\"langchain\",\"schema\",\"SystemMessage\"],\"kwargs\":{\"content\":\"Assistant is a large language model trained by OpenAI.\\n\\nAssistant is designed to be able to assist with a wide range of tasks, from answering simple questions to providing in-depth explanations and discussions on a wide range of topics. As a language model, Assistant is able to generate human-like text based on the input it receives, allowing it to engage in natural-sounding conversations and provide responses that are coherent and relevant to the topic at hand.\\n\\nAssistant is constantly learning and improving, and its capabilities are constantly evolving. It is able to process and understand large amounts of text, and can use this knowledge to provide accurate and informative responses to a wide range of questions. Additionally, Assistant is able to generate its own text based on the input it receives, allowing it to engage in discussions and provide explanations and descriptions on a wide range of topics.\\n\\nOverall, Assistant is a powerful system that can help with a wide range of tasks and provide valuable insights and information on a wide range of topics. Whether you need help with a specific question or just want to have a conversation about a particular topic, Assistant is here to assist. However, above all else, all responses must adhere to the format of RESPONSE FORMAT INSTRUCTIONS.\",\"additional_kwargs\":{}}},{\"lc\":1,\"type\":\"constructor\",\"id\":[\"langchain\",\"schema\",\"HumanMessage\"],\"kwargs\":{\"content\":\"TOOLS\\n------\\nAssistant can ask the user to use tools to look up information that may be helpful in answering the users original question. The tools the human can use are:\\n\\nesql-language-knowledge-base: Call this for knowledge on how to build an ESQL query, or answer questions about the ES|QL query language.\\n\\nRESPONSE FORMAT INSTRUCTIONS\\n----------------------------\\n\\nOutput a JSON markdown code snippet containing a valid JSON object in one of two formats:\\n\\n**Option 1:**\\nUse this if you want the human to use a tool.\\nMarkdown code snippet formatted in the following schema:\\n\\n```json\\n{\\n \\\"action\\\": string, // The action to take. Must be one of [esql-language-knowledge-base]\\n \\\"action_input\\\": string // The input to the action. May be a stringified object.\\n}\\n```\\n\\n**Option #2:**\\nUse this if you want to respond directly and conversationally to the human. Markdown code snippet formatted in the following schema:\\n\\n```json\\n{\\n \\\"action\\\": \\\"Final Answer\\\",\\n \\\"action_input\\\": string // You should put what you want to return to use here and make sure to use valid json newline characters.\\n}\\n```\\n\\nFor both options, remember to always include the surrounding markdown code snippet delimiters (begin with \\\"```json\\\" and end with \\\"```\\\")!\\n\\n\\nUSER'S INPUT\\n--------------------\\nHere is the user's input (remember to respond with a markdown code snippet of a json blob with a single action, and NOTHING else):\\n\\n\\n\\n\\n\\nFrom employees, I want to see the 5 earliest employees (hire_date), I want to display only the month and the year that they were hired in and their employee number (emp_no). Format the date as e.g. \\\"September 2019\\\". Only show the query\",\"additional_kwargs\":{}}}]" ] } [llm/end] [1:chain:AgentExecutor > 2:chain:LLMChain > 3:llm:ActionsClientLlm] [3.08s] Exiting LLM run with output: { "generations": [ [ { "text": "```json\n{\n \"action\": \"esql-language-knowledge-base\",\n \"action_input\": \"Display the 'emp_no', month and year of the 5 earliest employees by 'hire_date'. Format the date as 'Month Year'.\"\n}\n```" } ] ] } [chain/end] [1:chain:AgentExecutor > 2:chain:LLMChain] [3.09s] Exiting Chain run with output: { "text": "```json\n{\n \"action\": \"esql-language-knowledge-base\",\n \"action_input\": \"Display the 'emp_no', month and year of the 5 earliest employees by 'hire_date'. Format the date as 'Month Year'.\"\n}\n```" } [agent/action] [1:chain:AgentExecutor] Agent selected action: { "tool": "esql-language-knowledge-base", "toolInput": "Display the 'emp_no', month and year of the 5 earliest employees by 'hire_date'. Format the date as 'Month Year'.", "log": "```json\n{\n \"action\": \"esql-language-knowledge-base\",\n \"action_input\": \"Display the 'emp_no', month and year of the 5 earliest employees by 'hire_date'. Format the date as 'Month Year'.\"\n}\n```" } [tool/start] [1:chain:AgentExecutor > 4:tool:ChainTool] Entering Tool run with input: "Display the 'emp_no', month and year of the 5 earliest employees by 'hire_date'. Format the date as 'Month Year'." [chain/start] [1:chain:AgentExecutor > 4:tool:ChainTool > 5:chain:RetrievalQAChain] Entering Chain run with input: { "query": "Display the 'emp_no', month and year of the 5 earliest employees by 'hire_date'. Format the date as 'Month Year'." } [retriever/start] [1:chain:AgentExecutor > 4:tool:ChainTool > 5:chain:RetrievalQAChain > 6:retriever:VectorStoreRetriever] Entering Retriever run with input: { "query": "Display the 'emp_no', month and year of the 5 earliest employees by 'hire_date'. Format the date as 'Month Year'." } [retriever/end] [1:chain:AgentExecutor > 4:tool:ChainTool > 5:chain:RetrievalQAChain > 6:retriever:VectorStoreRetriever] [115ms] Exiting Retriever run with output: { "documents": [ { "pageContent": "[[esql-date_format]]\n=== `DATE_FORMAT`\nReturns a string representation of a date in the provided format. If no format\nis specified, the `yyyy-MM-dd'T'HH:mm:ss.SSSZ` format is used.\n\n[source,esql]\n----\nFROM employees\n| KEEP first_name, last_name, hire_date\n| EVAL hired = DATE_FORMAT(hire_date, \"YYYY-MM-dd\")\n----\n", "metadata": { "source": "/Users/andrew.goldstein/Projects/forks/spong/kibana/x-pack/plugins/elastic_assistant/server/knowledge_base/esql/docs/functions/date_format.asciidoc" } }, { "pageContent": "[[esql-date_trunc]]\n=== `DATE_TRUNC`\nRounds down a date to the closest interval. Intervals can be expressed using the\n<>.\n\n[source,esql]\n----\nFROM employees\n| EVAL year_hired = DATE_TRUNC(1 year, hire_date)\n| STATS count(emp_no) BY year_hired\n| SORT year_hired\n----\n", "metadata": { "source": "/Users/andrew.goldstein/Projects/forks/spong/kibana/x-pack/plugins/elastic_assistant/server/knowledge_base/esql/docs/functions/date_trunc.asciidoc" } }, { "pageContent": "[[esql-from]]\n=== `FROM`\n\nThe `FROM` source command returns a table with up to 10,000 documents from a\ndata stream, index, or alias. Each row in the resulting table represents a\ndocument. Each column corresponds to a field, and can be accessed by the name\nof that field.\n\n[source,esql]\n----\nFROM employees\n----\n\nYou can use <> to refer to indices, aliases\nand data streams. This can be useful for time series data, for example to access\ntoday's index:\n\n[source,esql]\n----\nFROM \n----\n\nUse comma-separated lists or wildcards to query multiple data streams, indices,\nor aliases:\n\n[source,esql]\n----\nFROM employees-00001,employees-*\n----\n", "metadata": { "source": "/Users/andrew.goldstein/Projects/forks/spong/kibana/x-pack/plugins/elastic_assistant/server/knowledge_base/esql/docs/source_commands/from.asciidoc" } }, { "pageContent": "[[esql-where]]\n=== `WHERE`\n\nUse `WHERE` to produce a table that contains all the rows from the input table\nfor which the provided condition evaluates to `true`:\n\n[source,esql]\n----\ninclude::{esql-specs}/docs.csv-spec[tag=where]\n----\n\nWhich, if `still_hired` is a boolean field, can be simplified to:\n\n[source,esql]\n----\ninclude::{esql-specs}/docs.csv-spec[tag=whereBoolean]\n----\n\n[discrete]\n==== Operators\n\nRefer to <> for an overview of the supported operators.\n\n[discrete]\n==== Functions\n`WHERE` supports various functions for calculating values. Refer to\n<> for more information.\n\n[source,esql]\n----\ninclude::{esql-specs}/docs.csv-spec[tag=whereFunction]\n----\n", "metadata": { "source": "/Users/andrew.goldstein/Projects/forks/spong/kibana/x-pack/plugins/elastic_assistant/server/knowledge_base/esql/docs/processing_commands/where.asciidoc" } } ] } [chain/start] [1:chain:AgentExecutor > 4:tool:ChainTool > 5:chain:RetrievalQAChain > 7:chain:StuffDocumentsChain] Entering Chain run with input: { "question": "Display the 'emp_no', month and year of the 5 earliest employees by 'hire_date'. Format the date as 'Month Year'.", "input_documents": [ { "pageContent": "[[esql-date_format]]\n=== `DATE_FORMAT`\nReturns a string representation of a date in the provided format. If no format\nis specified, the `yyyy-MM-dd'T'HH:mm:ss.SSSZ` format is used.\n\n[source,esql]\n----\nFROM employees\n| KEEP first_name, last_name, hire_date\n| EVAL hired = DATE_FORMAT(hire_date, \"YYYY-MM-dd\")\n----\n", "metadata": { "source": "/Users/andrew.goldstein/Projects/forks/spong/kibana/x-pack/plugins/elastic_assistant/server/knowledge_base/esql/docs/functions/date_format.asciidoc" } }, { "pageContent": "[[esql-date_trunc]]\n=== `DATE_TRUNC`\nRounds down a date to the closest interval. Intervals can be expressed using the\n<>.\n\n[source,esql]\n----\nFROM employees\n| EVAL year_hired = DATE_TRUNC(1 year, hire_date)\n| STATS count(emp_no) BY year_hired\n| SORT year_hired\n----\n", "metadata": { "source": "/Users/andrew.goldstein/Projects/forks/spong/kibana/x-pack/plugins/elastic_assistant/server/knowledge_base/esql/docs/functions/date_trunc.asciidoc" } }, { "pageContent": "[[esql-from]]\n=== `FROM`\n\nThe `FROM` source command returns a table with up to 10,000 documents from a\ndata stream, index, or alias. Each row in the resulting table represents a\ndocument. Each column corresponds to a field, and can be accessed by the name\nof that field.\n\n[source,esql]\n----\nFROM employees\n----\n\nYou can use <> to refer to indices, aliases\nand data streams. This can be useful for time series data, for example to access\ntoday's index:\n\n[source,esql]\n----\nFROM \n----\n\nUse comma-separated lists or wildcards to query multiple data streams, indices,\nor aliases:\n\n[source,esql]\n----\nFROM employees-00001,employees-*\n----\n", "metadata": { "source": "/Users/andrew.goldstein/Projects/forks/spong/kibana/x-pack/plugins/elastic_assistant/server/knowledge_base/esql/docs/source_commands/from.asciidoc" } }, { "pageContent": "[[esql-where]]\n=== `WHERE`\n\nUse `WHERE` to produce a table that contains all the rows from the input table\nfor which the provided condition evaluates to `true`:\n\n[source,esql]\n----\ninclude::{esql-specs}/docs.csv-spec[tag=where]\n----\n\nWhich, if `still_hired` is a boolean field, can be simplified to:\n\n[source,esql]\n----\ninclude::{esql-specs}/docs.csv-spec[tag=whereBoolean]\n----\n\n[discrete]\n==== Operators\n\nRefer to <> for an overview of the supported operators.\n\n[discrete]\n==== Functions\n`WHERE` supports various functions for calculating values. Refer to\n<> for more information.\n\n[source,esql]\n----\ninclude::{esql-specs}/docs.csv-spec[tag=whereFunction]\n----\n", "metadata": { "source": "/Users/andrew.goldstein/Projects/forks/spong/kibana/x-pack/plugins/elastic_assistant/server/knowledge_base/esql/docs/processing_commands/where.asciidoc" } } ], "query": "Display the 'emp_no', month and year of the 5 earliest employees by 'hire_date'. Format the date as 'Month Year'." } [chain/start] [1:chain:AgentExecutor > 4:tool:ChainTool > 5:chain:RetrievalQAChain > 7:chain:StuffDocumentsChain > 8:chain:LLMChain] Entering Chain run with input: { "question": "Display the 'emp_no', month and year of the 5 earliest employees by 'hire_date'. Format the date as 'Month Year'.", "query": "Display the 'emp_no', month and year of the 5 earliest employees by 'hire_date'. Format the date as 'Month Year'.", "context": "[[esql-date_format]]\n=== `DATE_FORMAT`\nReturns a string representation of a date in the provided format. If no format\nis specified, the `yyyy-MM-dd'T'HH:mm:ss.SSSZ` format is used.\n\n[source,esql]\n----\nFROM employees\n| KEEP first_name, last_name, hire_date\n| EVAL hired = DATE_FORMAT(hire_date, \"YYYY-MM-dd\")\n----\n\n\n[[esql-date_trunc]]\n=== `DATE_TRUNC`\nRounds down a date to the closest interval. Intervals can be expressed using the\n<>.\n\n[source,esql]\n----\nFROM employees\n| EVAL year_hired = DATE_TRUNC(1 year, hire_date)\n| STATS count(emp_no) BY year_hired\n| SORT year_hired\n----\n\n\n[[esql-from]]\n=== `FROM`\n\nThe `FROM` source command returns a table with up to 10,000 documents from a\ndata stream, index, or alias. Each row in the resulting table represents a\ndocument. Each column corresponds to a field, and can be accessed by the name\nof that field.\n\n[source,esql]\n----\nFROM employees\n----\n\nYou can use <> to refer to indices, aliases\nand data streams. This can be useful for time series data, for example to access\ntoday's index:\n\n[source,esql]\n----\nFROM \n----\n\nUse comma-separated lists or wildcards to query multiple data streams, indices,\nor aliases:\n\n[source,esql]\n----\nFROM employees-00001,employees-*\n----\n\n\n[[esql-where]]\n=== `WHERE`\n\nUse `WHERE` to produce a table that contains all the rows from the input table\nfor which the provided condition evaluates to `true`:\n\n[source,esql]\n----\ninclude::{esql-specs}/docs.csv-spec[tag=where]\n----\n\nWhich, if `still_hired` is a boolean field, can be simplified to:\n\n[source,esql]\n----\ninclude::{esql-specs}/docs.csv-spec[tag=whereBoolean]\n----\n\n[discrete]\n==== Operators\n\nRefer to <> for an overview of the supported operators.\n\n[discrete]\n==== Functions\n`WHERE` supports various functions for calculating values. Refer to\n<> for more information.\n\n[source,esql]\n----\ninclude::{esql-specs}/docs.csv-spec[tag=whereFunction]\n----\n" } [llm/start] [1:chain:AgentExecutor > 4:tool:ChainTool > 5:chain:RetrievalQAChain > 7:chain:StuffDocumentsChain > 8:chain:LLMChain > 9:llm:ActionsClientLlm] Entering LLM run with input: { "prompts": [ "Use the following pieces of context to answer the question at the end. If you don't know the answer, just say that you don't know, don't try to make up an answer.\n\n[[esql-date_format]]\n=== `DATE_FORMAT`\nReturns a string representation of a date in the provided format. If no format\nis specified, the `yyyy-MM-dd'T'HH:mm:ss.SSSZ` format is used.\n\n[source,esql]\n----\nFROM employees\n| KEEP first_name, last_name, hire_date\n| EVAL hired = DATE_FORMAT(hire_date, \"YYYY-MM-dd\")\n----\n\n\n[[esql-date_trunc]]\n=== `DATE_TRUNC`\nRounds down a date to the closest interval. Intervals can be expressed using the\n<>.\n\n[source,esql]\n----\nFROM employees\n| EVAL year_hired = DATE_TRUNC(1 year, hire_date)\n| STATS count(emp_no) BY year_hired\n| SORT year_hired\n----\n\n\n[[esql-from]]\n=== `FROM`\n\nThe `FROM` source command returns a table with up to 10,000 documents from a\ndata stream, index, or alias. Each row in the resulting table represents a\ndocument. Each column corresponds to a field, and can be accessed by the name\nof that field.\n\n[source,esql]\n----\nFROM employees\n----\n\nYou can use <> to refer to indices, aliases\nand data streams. This can be useful for time series data, for example to access\ntoday's index:\n\n[source,esql]\n----\nFROM \n----\n\nUse comma-separated lists or wildcards to query multiple data streams, indices,\nor aliases:\n\n[source,esql]\n----\nFROM employees-00001,employees-*\n----\n\n\n[[esql-where]]\n=== `WHERE`\n\nUse `WHERE` to produce a table that contains all the rows from the input table\nfor which the provided condition evaluates to `true`:\n\n[source,esql]\n----\ninclude::{esql-specs}/docs.csv-spec[tag=where]\n----\n\nWhich, if `still_hired` is a boolean field, can be simplified to:\n\n[source,esql]\n----\ninclude::{esql-specs}/docs.csv-spec[tag=whereBoolean]\n----\n\n[discrete]\n==== Operators\n\nRefer to <> for an overview of the supported operators.\n\n[discrete]\n==== Functions\n`WHERE` supports various functions for calculating values. Refer to\n<> for more information.\n\n[source,esql]\n----\ninclude::{esql-specs}/docs.csv-spec[tag=whereFunction]\n----\n\n\nQuestion: Display the 'emp_no', month and year of the 5 earliest employees by 'hire_date'. Format the date as 'Month Year'.\nHelpful Answer:" ] } [llm/end] [1:chain:AgentExecutor > 4:tool:ChainTool > 5:chain:RetrievalQAChain > 7:chain:StuffDocumentsChain > 8:chain:LLMChain > 9:llm:ActionsClientLlm] [2.23s] Exiting LLM run with output: { "generations": [ [ { "text": "FROM employees\n| KEEP emp_no, hire_date\n| EVAL month_year = DATE_FORMAT(hire_date, \"MMMM YYYY\")\n| SORT hire_date\n| LIMIT 5" } ] ] } [chain/end] [1:chain:AgentExecutor > 4:tool:ChainTool > 5:chain:RetrievalQAChain > 7:chain:StuffDocumentsChain > 8:chain:LLMChain] [2.23s] Exiting Chain run with output: { "text": "FROM employees\n| KEEP emp_no, hire_date\n| EVAL month_year = DATE_FORMAT(hire_date, \"MMMM YYYY\")\n| SORT hire_date\n| LIMIT 5" } [chain/end] [1:chain:AgentExecutor > 4:tool:ChainTool > 5:chain:RetrievalQAChain > 7:chain:StuffDocumentsChain] [2.23s] Exiting Chain run with output: { "text": "FROM employees\n| KEEP emp_no, hire_date\n| EVAL month_year = DATE_FORMAT(hire_date, \"MMMM YYYY\")\n| SORT hire_date\n| LIMIT 5" } [chain/end] [1:chain:AgentExecutor > 4:tool:ChainTool > 5:chain:RetrievalQAChain] [2.35s] Exiting Chain run with output: { "text": "FROM employees\n| KEEP emp_no, hire_date\n| EVAL month_year = DATE_FORMAT(hire_date, \"MMMM YYYY\")\n| SORT hire_date\n| LIMIT 5" } [tool/end] [1:chain:AgentExecutor > 4:tool:ChainTool] [2.35s] Exiting Tool run with output: "FROM employees | KEEP emp_no, hire_date | EVAL month_year = DATE_FORMAT(hire_date, "MMMM YYYY") | SORT hire_date | LIMIT 5" [chain/start] [1:chain:AgentExecutor > 10:chain:LLMChain] Entering Chain run with input: { "input": "\n\n\n\nFrom employees, I want to see the 5 earliest employees (hire_date), I want to display only the month and the year that they were hired in and their employee number (emp_no). Format the date as e.g. \"September 2019\". Only show the query", "chat_history": [], "agent_scratchpad": [ { "lc": 1, "type": "constructor", "id": [ "langchain", "schema", "AIMessage" ], "kwargs": { "content": "```json\n{\n \"action\": \"esql-language-knowledge-base\",\n \"action_input\": \"Display the 'emp_no', month and year of the 5 earliest employees by 'hire_date'. Format the date as 'Month Year'.\"\n}\n```", "additional_kwargs": {} } }, { "lc": 1, "type": "constructor", "id": [ "langchain", "schema", "HumanMessage" ], "kwargs": { "content": "TOOL RESPONSE:\n---------------------\nFROM employees\n| KEEP emp_no, hire_date\n| EVAL month_year = DATE_FORMAT(hire_date, \"MMMM YYYY\")\n| SORT hire_date\n| LIMIT 5\n\nUSER'S INPUT\n--------------------\n\nOkay, so what is the response to my last comment? If using information obtained from the tools you must mention it explicitly without mentioning the tool names - I have forgotten all TOOL RESPONSES! Remember to respond with a markdown code snippet of a json blob with a single action, and NOTHING else.", "additional_kwargs": {} } } ], "stop": [ "Observation:" ] } [llm/start] [1:chain:AgentExecutor > 10:chain:LLMChain > 11:llm:ActionsClientLlm] Entering LLM run with input: { "prompts": [ "[{\"lc\":1,\"type\":\"constructor\",\"id\":[\"langchain\",\"schema\",\"SystemMessage\"],\"kwargs\":{\"content\":\"Assistant is a large language model trained by OpenAI.\\n\\nAssistant is designed to be able to assist with a wide range of tasks, from answering simple questions to providing in-depth explanations and discussions on a wide range of topics. As a language model, Assistant is able to generate human-like text based on the input it receives, allowing it to engage in natural-sounding conversations and provide responses that are coherent and relevant to the topic at hand.\\n\\nAssistant is constantly learning and improving, and its capabilities are constantly evolving. It is able to process and understand large amounts of text, and can use this knowledge to provide accurate and informative responses to a wide range of questions. Additionally, Assistant is able to generate its own text based on the input it receives, allowing it to engage in discussions and provide explanations and descriptions on a wide range of topics.\\n\\nOverall, Assistant is a powerful system that can help with a wide range of tasks and provide valuable insights and information on a wide range of topics. Whether you need help with a specific question or just want to have a conversation about a particular topic, Assistant is here to assist. However, above all else, all responses must adhere to the format of RESPONSE FORMAT INSTRUCTIONS.\",\"additional_kwargs\":{}}},{\"lc\":1,\"type\":\"constructor\",\"id\":[\"langchain\",\"schema\",\"HumanMessage\"],\"kwargs\":{\"content\":\"TOOLS\\n------\\nAssistant can ask the user to use tools to look up information that may be helpful in answering the users original question. The tools the human can use are:\\n\\nesql-language-knowledge-base: Call this for knowledge on how to build an ESQL query, or answer questions about the ES|QL query language.\\n\\nRESPONSE FORMAT INSTRUCTIONS\\n----------------------------\\n\\nOutput a JSON markdown code snippet containing a valid JSON object in one of two formats:\\n\\n**Option 1:**\\nUse this if you want the human to use a tool.\\nMarkdown code snippet formatted in the following schema:\\n\\n```json\\n{\\n \\\"action\\\": string, // The action to take. Must be one of [esql-language-knowledge-base]\\n \\\"action_input\\\": string // The input to the action. May be a stringified object.\\n}\\n```\\n\\n**Option #2:**\\nUse this if you want to respond directly and conversationally to the human. Markdown code snippet formatted in the following schema:\\n\\n```json\\n{\\n \\\"action\\\": \\\"Final Answer\\\",\\n \\\"action_input\\\": string // You should put what you want to return to use here and make sure to use valid json newline characters.\\n}\\n```\\n\\nFor both options, remember to always include the surrounding markdown code snippet delimiters (begin with \\\"```json\\\" and end with \\\"```\\\")!\\n\\n\\nUSER'S INPUT\\n--------------------\\nHere is the user's input (remember to respond with a markdown code snippet of a json blob with a single action, and NOTHING else):\\n\\n\\n\\n\\n\\nFrom employees, I want to see the 5 earliest employees (hire_date), I want to display only the month and the year that they were hired in and their employee number (emp_no). Format the date as e.g. \\\"September 2019\\\". Only show the query\",\"additional_kwargs\":{}}},{\"lc\":1,\"type\":\"constructor\",\"id\":[\"langchain\",\"schema\",\"AIMessage\"],\"kwargs\":{\"content\":\"```json\\n{\\n \\\"action\\\": \\\"esql-language-knowledge-base\\\",\\n \\\"action_input\\\": \\\"Display the 'emp_no', month and year of the 5 earliest employees by 'hire_date'. Format the date as 'Month Year'.\\\"\\n}\\n```\",\"additional_kwargs\":{}}},{\"lc\":1,\"type\":\"constructor\",\"id\":[\"langchain\",\"schema\",\"HumanMessage\"],\"kwargs\":{\"content\":\"TOOL RESPONSE:\\n---------------------\\nFROM employees\\n| KEEP emp_no, hire_date\\n| EVAL month_year = DATE_FORMAT(hire_date, \\\"MMMM YYYY\\\")\\n| SORT hire_date\\n| LIMIT 5\\n\\nUSER'S INPUT\\n--------------------\\n\\nOkay, so what is the response to my last comment? If using information obtained from the tools you must mention it explicitly without mentioning the tool names - I have forgotten all TOOL RESPONSES! Remember to respond with a markdown code snippet of a json blob with a single action, and NOTHING else.\",\"additional_kwargs\":{}}}]" ] } [llm/end] [1:chain:AgentExecutor > 10:chain:LLMChain > 11:llm:ActionsClientLlm] [6.47s] Exiting LLM run with output: { "generations": [ [ { "text": "```json\n{\n \"action\": \"Final Answer\",\n \"action_input\": \"Here is the query to get the employee number and the formatted hire date for the 5 earliest employees by hire_date:\\n\\nFROM employees\\n| KEEP emp_no, hire_date\\n| EVAL month_year = DATE_FORMAT(hire_date, \\\"MMMM YYYY\\\")\\n| SORT hire_date\\n| LIMIT 5\"\n}\n```" } ] ] } [chain/end] [1:chain:AgentExecutor > 10:chain:LLMChain] [6.47s] Exiting Chain run with output: { "text": "```json\n{\n \"action\": \"Final Answer\",\n \"action_input\": \"Here is the query to get the employee number and the formatted hire date for the 5 earliest employees by hire_date:\\n\\nFROM employees\\n| KEEP emp_no, hire_date\\n| EVAL month_year = DATE_FORMAT(hire_date, \\\"MMMM YYYY\\\")\\n| SORT hire_date\\n| LIMIT 5\"\n}\n```" } [chain/end] [1:chain:AgentExecutor] [11.91s] Exiting Chain run with output: { "output": "Here is the query to get the employee number and the formatted hire date for the 5 earliest employees by hire_date:\n\nFROM employees\n| KEEP emp_no, hire_date\n| EVAL month_year = DATE_FORMAT(hire_date, \"MMMM YYYY\")\n| SORT hire_date\n| LIMIT 5" } ``` --- .../impl/assistant/api.test.tsx | 84 +++++++++++++++++++ .../impl/assistant/api.tsx | 4 +- .../impl/assistant/helpers.test.ts | 43 +++++++++- .../impl/assistant/helpers.ts | 21 +++++ .../execute_custom_llm_chain/index.test.ts | 29 ++++--- .../execute_custom_llm_chain/index.ts | 39 +++++---- .../post_actions_connector_execute.test.ts | 2 +- .../routes/post_actions_connector_execute.ts | 4 +- 8 files changed, 195 insertions(+), 31 deletions(-) diff --git a/x-pack/packages/kbn-elastic-assistant/impl/assistant/api.test.tsx b/x-pack/packages/kbn-elastic-assistant/impl/assistant/api.test.tsx index 65b8183b60a0b..2f46e99d12b07 100644 --- a/x-pack/packages/kbn-elastic-assistant/impl/assistant/api.test.tsx +++ b/x-pack/packages/kbn-elastic-assistant/impl/assistant/api.test.tsx @@ -126,4 +126,88 @@ describe('fetchConnectorExecuteAction', () => { expect(result).toBe('Test response'); }); + + it('returns the value of the action_input property when assistantLangChain is true, and `content` has properly prefixed and suffixed JSON with the action_input property', async () => { + const content = '```json\n{"action_input": "value from action_input"}\n```'; + + (mockHttp.fetch as jest.Mock).mockResolvedValue({ + status: 'ok', + data: { + choices: [ + { + message: { + content, + }, + }, + ], + }, + }); + + const testProps: FetchConnectorExecuteAction = { + assistantLangChain: true, // <-- requires response parsing + http: mockHttp, + messages, + apiConfig, + }; + + const result = await fetchConnectorExecuteAction(testProps); + + expect(result).toBe('value from action_input'); + }); + + it('returns the original content when assistantLangChain is true, and `content` has properly formatted JSON WITHOUT the action_input property', async () => { + const content = '```json\n{"some_key": "some value"}\n```'; + + (mockHttp.fetch as jest.Mock).mockResolvedValue({ + status: 'ok', + data: { + choices: [ + { + message: { + content, + }, + }, + ], + }, + }); + + const testProps: FetchConnectorExecuteAction = { + assistantLangChain: true, // <-- requires response parsing + http: mockHttp, + messages, + apiConfig, + }; + + const result = await fetchConnectorExecuteAction(testProps); + + expect(result).toBe(content); + }); + + it('returns the original when assistantLangChain is true, and `content` is not JSON', async () => { + const content = 'plain text content'; + + (mockHttp.fetch as jest.Mock).mockResolvedValue({ + status: 'ok', + data: { + choices: [ + { + message: { + content, + }, + }, + ], + }, + }); + + const testProps: FetchConnectorExecuteAction = { + assistantLangChain: true, // <-- requires response parsing + http: mockHttp, + messages, + apiConfig, + }; + + const result = await fetchConnectorExecuteAction(testProps); + + expect(result).toBe(content); + }); }); diff --git a/x-pack/packages/kbn-elastic-assistant/impl/assistant/api.tsx b/x-pack/packages/kbn-elastic-assistant/impl/assistant/api.tsx index 511b5aa585af0..6d3452b6f7880 100644 --- a/x-pack/packages/kbn-elastic-assistant/impl/assistant/api.tsx +++ b/x-pack/packages/kbn-elastic-assistant/impl/assistant/api.tsx @@ -12,6 +12,7 @@ import { HttpSetup, IHttpFetchError } from '@kbn/core-http-browser'; import type { Conversation, Message } from '../assistant_context/types'; import { API_ERROR } from './translations'; import { MODEL_GPT_3_5_TURBO } from '../connectorland/models/model_selector/model_selector'; +import { getFormattedMessageContent } from './helpers'; export interface FetchConnectorExecuteAction { assistantLangChain: boolean; @@ -78,7 +79,8 @@ export const fetchConnectorExecuteAction = async ({ if (data.choices && data.choices.length > 0 && data.choices[0].message.content) { const result = data.choices[0].message.content.trim(); - return result; + + return assistantLangChain ? getFormattedMessageContent(result) : result; } else { return API_ERROR; } diff --git a/x-pack/packages/kbn-elastic-assistant/impl/assistant/helpers.test.ts b/x-pack/packages/kbn-elastic-assistant/impl/assistant/helpers.test.ts index 69bed887e730e..f2b89a07c319e 100644 --- a/x-pack/packages/kbn-elastic-assistant/impl/assistant/helpers.test.ts +++ b/x-pack/packages/kbn-elastic-assistant/impl/assistant/helpers.test.ts @@ -5,7 +5,11 @@ * 2.0. */ -import { getDefaultConnector, getBlockBotConversation } from './helpers'; +import { + getBlockBotConversation, + getDefaultConnector, + getFormattedMessageContent, +} from './helpers'; import { enterpriseMessaging } from './use_conversation/sample_conversations'; import { ActionConnector } from '@kbn/triggers-actions-ui-plugin/public'; @@ -190,4 +194,41 @@ describe('getBlockBotConversation', () => { expect(result).toBeUndefined(); }); }); + + describe('getFormattedMessageContent', () => { + it('returns the value of the action_input property when `content` has properly prefixed and suffixed JSON with the action_input property', () => { + const content = '```json\n{"action_input": "value from action_input"}\n```'; + + expect(getFormattedMessageContent(content)).toBe('value from action_input'); + }); + + it('returns the original content when `content` has properly formatted JSON WITHOUT the action_input property', () => { + const content = '```json\n{"some_key": "some value"}\n```'; + expect(getFormattedMessageContent(content)).toBe(content); + }); + + it('returns the original content when `content` has improperly formatted JSON', () => { + const content = '```json\n{"action_input": "value from action_input",}\n```'; // <-- the trailing comma makes it invalid + + expect(getFormattedMessageContent(content)).toBe(content); + }); + + it('returns the original content when `content` is missing the prefix', () => { + const content = '{"action_input": "value from action_input"}\n```'; // <-- missing prefix + + expect(getFormattedMessageContent(content)).toBe(content); + }); + + it('returns the original content when `content` is missing the suffix', () => { + const content = '```json\n{"action_input": "value from action_input"}'; // <-- missing suffix + + expect(getFormattedMessageContent(content)).toBe(content); + }); + + it('returns the original content when `content` does NOT contain a JSON string', () => { + const content = 'plain text content'; + + expect(getFormattedMessageContent(content)).toBe(content); + }); + }); }); diff --git a/x-pack/packages/kbn-elastic-assistant/impl/assistant/helpers.ts b/x-pack/packages/kbn-elastic-assistant/impl/assistant/helpers.ts index b01c9001e8319..2b2c5b76851f7 100644 --- a/x-pack/packages/kbn-elastic-assistant/impl/assistant/helpers.ts +++ b/x-pack/packages/kbn-elastic-assistant/impl/assistant/helpers.ts @@ -59,3 +59,24 @@ export const getDefaultConnector = ( connectors: Array, Record>> | undefined ): ActionConnector, Record> | undefined => connectors?.length === 1 ? connectors[0] : undefined; + +/** + * When `content` is a JSON string, prefixed with "```json\n" + * and suffixed with "\n```", this function will attempt to parse it and return + * the `action_input` property if it exists. + */ +export const getFormattedMessageContent = (content: string): string => { + const formattedContentMatch = content.match(/```json\n([\s\S]+)\n```/); + + if (formattedContentMatch) { + try { + const parsedContent = JSON.parse(formattedContentMatch[1]); + + return parsedContent.action_input ?? content; + } catch { + // we don't want to throw an error here, so we'll fall back to the original content + } + } + + return content; +}; diff --git a/x-pack/plugins/elastic_assistant/server/lib/langchain/execute_custom_llm_chain/index.test.ts b/x-pack/plugins/elastic_assistant/server/lib/langchain/execute_custom_llm_chain/index.test.ts index be1adbc2e1ce4..67fb3859b9943 100644 --- a/x-pack/plugins/elastic_assistant/server/lib/langchain/execute_custom_llm_chain/index.test.ts +++ b/x-pack/plugins/elastic_assistant/server/lib/langchain/execute_custom_llm_chain/index.test.ts @@ -12,7 +12,7 @@ import { ResponseBody } from '../helpers'; import { ActionsClientLlm } from '../llm/actions_client_llm'; import { mockActionResultData } from '../../../__mocks__/action_result_data'; import { langChainMessages } from '../../../__mocks__/lang_chain_messages'; -import { executeCustomLlmChain } from '.'; +import { callAgentExecutor } from '.'; import { loggerMock } from '@kbn/logging-mocks'; import { elasticsearchServiceMock } from '@kbn/core-elasticsearch-server-mocks'; @@ -23,11 +23,18 @@ const mockConversationChain = { }; jest.mock('langchain/chains', () => ({ - ConversationalRetrievalQAChain: { + RetrievalQAChain: { fromLLM: jest.fn().mockImplementation(() => mockConversationChain), }, })); +const mockCall = jest.fn(); +jest.mock('langchain/agents', () => ({ + initializeAgentExecutorWithOptions: jest.fn().mockImplementation(() => ({ + call: mockCall, + })), +})); + const mockConnectorId = 'mock-connector-id'; // eslint-disable-next-line @typescript-eslint/no-explicit-any @@ -42,7 +49,7 @@ const mockActions: ActionsPluginStart = {} as ActionsPluginStart; const mockLogger = loggerMock.create(); const esClientMock = elasticsearchServiceMock.createScopedClusterClient().asCurrentUser; -describe('executeCustomLlmChain', () => { +describe('callAgentExecutor', () => { beforeEach(() => { jest.clearAllMocks(); @@ -52,7 +59,7 @@ describe('executeCustomLlmChain', () => { }); it('creates an instance of ActionsClientLlm with the expected context from the request', async () => { - await executeCustomLlmChain({ + await callAgentExecutor({ actions: mockActions, connectorId: mockConnectorId, esClient: esClientMock, @@ -70,7 +77,7 @@ describe('executeCustomLlmChain', () => { }); it('kicks off the chain with (only) the last message', async () => { - await executeCustomLlmChain({ + await callAgentExecutor({ actions: mockActions, connectorId: mockConnectorId, esClient: esClientMock, @@ -79,15 +86,15 @@ describe('executeCustomLlmChain', () => { request: mockRequest, }); - expect(mockConversationChain.call).toHaveBeenCalledWith({ - question: '\n\nDo you know my name?', + expect(mockCall).toHaveBeenCalledWith({ + input: '\n\nDo you know my name?', }); }); it('kicks off the chain with the expected message when langChainMessages has only one entry', async () => { const onlyOneMessage = [langChainMessages[0]]; - await executeCustomLlmChain({ + await callAgentExecutor({ actions: mockActions, connectorId: mockConnectorId, esClient: esClientMock, @@ -96,13 +103,13 @@ describe('executeCustomLlmChain', () => { request: mockRequest, }); - expect(mockConversationChain.call).toHaveBeenCalledWith({ - question: 'What is my name?', + expect(mockCall).toHaveBeenCalledWith({ + input: 'What is my name?', }); }); it('returns the expected response body', async () => { - const result: ResponseBody = await executeCustomLlmChain({ + const result: ResponseBody = await callAgentExecutor({ actions: mockActions, connectorId: mockConnectorId, esClient: esClientMock, diff --git a/x-pack/plugins/elastic_assistant/server/lib/langchain/execute_custom_llm_chain/index.ts b/x-pack/plugins/elastic_assistant/server/lib/langchain/execute_custom_llm_chain/index.ts index 5a65b1589b21e..b6a768ad69598 100644 --- a/x-pack/plugins/elastic_assistant/server/lib/langchain/execute_custom_llm_chain/index.ts +++ b/x-pack/plugins/elastic_assistant/server/lib/langchain/execute_custom_llm_chain/index.ts @@ -7,16 +7,18 @@ import { ElasticsearchClient, KibanaRequest, Logger } from '@kbn/core/server'; import type { PluginStartContract as ActionsPluginStart } from '@kbn/actions-plugin/server'; +import { initializeAgentExecutorWithOptions } from 'langchain/agents'; +import { RetrievalQAChain } from 'langchain/chains'; import { BufferMemory, ChatMessageHistory } from 'langchain/memory'; import { BaseMessage } from 'langchain/schema'; +import { ChainTool, Tool } from 'langchain/tools'; -import { ConversationalRetrievalQAChain } from 'langchain/chains'; +import { ElasticsearchStore } from '../elasticsearch_store/elasticsearch_store'; import { ResponseBody } from '../helpers'; import { ActionsClientLlm } from '../llm/actions_client_llm'; -import { ElasticsearchStore } from '../elasticsearch_store/elasticsearch_store'; import { KNOWLEDGE_BASE_INDEX_PATTERN } from '../../../routes/knowledge_base/constants'; -export const executeCustomLlmChain = async ({ +export const callAgentExecutor = async ({ actions, connectorId, esClient, @@ -34,31 +36,38 @@ export const executeCustomLlmChain = async ({ }): Promise => { const llm = new ActionsClientLlm({ actions, connectorId, request, logger }); - // Chat History Memory: in-memory memory, from client local storage, first message is the system prompt const pastMessages = langChainMessages.slice(0, -1); // all but the last message const latestMessage = langChainMessages.slice(-1); // the last message + const memory = new BufferMemory({ chatHistory: new ChatMessageHistory(pastMessages), - memoryKey: 'chat_history', + memoryKey: 'chat_history', // this is the key expected by https://github.com/langchain-ai/langchainjs/blob/a13a8969345b0f149c1ca4a120d63508b06c52a5/langchain/src/agents/initialize.ts#L166 + inputKey: 'input', + outputKey: 'output', + returnMessages: true, }); // ELSER backed ElasticsearchStore for Knowledge Base const esStore = new ElasticsearchStore(esClient, KNOWLEDGE_BASE_INDEX_PATTERN, logger); + const chain = RetrievalQAChain.fromLLM(llm, esStore.asRetriever()); + + const tools: Tool[] = [ + new ChainTool({ + name: 'esql-language-knowledge-base', + description: + 'Call this for knowledge on how to build an ESQL query, or answer questions about the ES|QL query language.', + chain, + }), + ]; - // Chain w/ chat history memory and knowledge base retriever - const chain = ConversationalRetrievalQAChain.fromLLM(llm, esStore.asRetriever(), { + const executor = await initializeAgentExecutorWithOptions(tools, llm, { + agentType: 'chat-conversational-react-description', memory, - // See `qaChainOptions` from https://js.langchain.com/docs/modules/chains/popular/chat_vector_db - qaChainOptions: { type: 'stuff' }, + verbose: false, }); - await chain.call({ question: latestMessage[0].content }); - // Chain w/ just knowledge base retriever - // const chain = RetrievalQAChain.fromLLM(llm, esStore.asRetriever()); - // await chain.call({ query: latestMessage[0].content }); + await executor.call({ input: latestMessage[0].content }); - // The assistant (on the client side) expects the same response returned - // from the actions framework, so we need to return the same shape of data: return { connector_id: connectorId, data: llm.getActionResultData(), // the response from the actions framework diff --git a/x-pack/plugins/elastic_assistant/server/routes/post_actions_connector_execute.test.ts b/x-pack/plugins/elastic_assistant/server/routes/post_actions_connector_execute.test.ts index 2e6709a6e33c2..57f2b25f5a65f 100644 --- a/x-pack/plugins/elastic_assistant/server/routes/post_actions_connector_execute.test.ts +++ b/x-pack/plugins/elastic_assistant/server/routes/post_actions_connector_execute.test.ts @@ -20,7 +20,7 @@ jest.mock('../lib/build_response', () => ({ })); jest.mock('../lib/langchain/execute_custom_llm_chain', () => ({ - executeCustomLlmChain: jest.fn().mockImplementation( + callAgentExecutor: jest.fn().mockImplementation( async ({ connectorId, }: { diff --git a/x-pack/plugins/elastic_assistant/server/routes/post_actions_connector_execute.ts b/x-pack/plugins/elastic_assistant/server/routes/post_actions_connector_execute.ts index 1043f68f0f9c1..bbb1c76e3e579 100644 --- a/x-pack/plugins/elastic_assistant/server/routes/post_actions_connector_execute.ts +++ b/x-pack/plugins/elastic_assistant/server/routes/post_actions_connector_execute.ts @@ -20,7 +20,7 @@ import { PostActionsConnectorExecutePathParams, } from '../schemas/post_actions_connector_execute'; import { ElasticAssistantRequestHandlerContext } from '../types'; -import { executeCustomLlmChain } from '../lib/langchain/execute_custom_llm_chain'; +import { callAgentExecutor } from '../lib/langchain/execute_custom_llm_chain'; export const postActionsConnectorExecuteRoute = ( router: IRouter @@ -53,7 +53,7 @@ export const postActionsConnectorExecuteRoute = ( // convert the assistant messages to LangChain messages: const langChainMessages = getLangChainMessages(assistantMessages); - const langChainResponseBody = await executeCustomLlmChain({ + const langChainResponseBody = await callAgentExecutor({ actions, connectorId, esClient,