@memtensor/memos-local-openclaw-plugin 0.1.3 → 0.1.5

package/src/ingest/providers/bedrock.ts

@@ -1,12 +1,221 @@
 import type { SummarizerConfig, Logger } from "../../types";
 
-const SYSTEM_PROMPT = `Summarize the text in ONE concise sentence (max 60 tokens). Preserve exact names, commands, error codes. No bullet points, no preamble — output only the sentence.`;
+const SYSTEM_PROMPT = `Summarize the text in ONE concise sentence (max 120 characters). IMPORTANT: Use the SAME language as the input text — if the input is Chinese, write Chinese; if English, write English. Preserve exact names, commands, error codes. No bullet points, no preamble — output only the sentence.`;
+
+const TASK_SUMMARY_PROMPT = `You create a DETAILED task summary from a multi-turn conversation. This summary will be the ONLY record of this conversation, so it must preserve ALL important information.
+
+CRITICAL LANGUAGE RULE: You MUST write in the SAME language as the user's messages. Chinese input → Chinese output. English input → English output. NEVER mix languages.
+
+Output EXACTLY this structure:
+
+📌 Title
+A short, descriptive title (10-30 characters). Like a chat group name.
+
+🎯 Goal
+One sentence: what the user wanted to accomplish.
+
+📋 Key Steps
+- Describe each meaningful step in detail
+- Include the ACTUAL content produced: code snippets, commands, config blocks, formulas, key paragraphs
+- For code: include the function signature and core logic (up to ~30 lines per block), use fenced code blocks
+- For configs: include the actual config values and structure
+- For lists/instructions: include the actual items, not just "provided a list"
+- Merge only truly trivial back-and-forth (like "ok" / "sure")
+- Do NOT over-summarize: "provided a function" is BAD; show the actual function
+
+✅ Result
+What was the final outcome? Include the final version of any code/config/content produced.
+
+💡 Key Details
+- Decisions made, trade-offs discussed, caveats noted, alternative approaches mentioned
+- Specific values: numbers, versions, thresholds, URLs, file paths, model names
+- Omit this section only if there truly are no noteworthy details
+
+RULES:
+- This summary is a KNOWLEDGE BASE ENTRY, not a brief note. Be thorough.
+- PRESERVE verbatim: code, commands, URLs, file paths, error messages, config values, version numbers, names, amounts
+- DISCARD only: greetings, filler, the assistant explaining what it will do before doing it
+- Replace secrets (API keys, tokens, passwords) with [REDACTED]
+- Target length: 30-50% of the original conversation length. Longer conversations need longer summaries.
+- Output summary only, no preamble.`;
+
+export async function summarizeTaskBedrock(
+  text: string,
+  cfg: SummarizerConfig,
+  log: Logger,
+): Promise<string> {
+  const model = cfg.model ?? "anthropic.claude-3-haiku-20240307-v1:0";
+  const endpoint = cfg.endpoint;
+  if (!endpoint) {
+    throw new Error("Bedrock task-summarizer requires 'endpoint'");
+  }
+
+  const url = `${endpoint}/model/${model}/converse`;
+  const headers: Record<string, string> = {
+    "Content-Type": "application/json",
+    ...cfg.headers,
+  };
+
+  const resp = await fetch(url, {
+    method: "POST",
+    headers,
+    body: JSON.stringify({
+      system: [{ text: TASK_SUMMARY_PROMPT }],
+      messages: [{ role: "user", content: [{ text }] }],
+      inferenceConfig: { temperature: cfg.temperature ?? 0.1, maxTokens: 4096 },
+    }),
+    signal: AbortSignal.timeout(cfg.timeoutMs ?? 60_000),
+  });
+
+  if (!resp.ok) {
+    const body = await resp.text();
+    throw new Error(`Bedrock task-summarize failed (${resp.status}): ${body}`);
+  }
+
+  const json = (await resp.json()) as { output: { message: { content: Array<{ text: string }> } } };
+  return json.output?.message?.content?.[0]?.text?.trim() ?? "";
+}
+
+const TOPIC_JUDGE_PROMPT = `You are a conversation topic boundary detector. Given a summary of the CURRENT conversation and a NEW user message, determine if the new message starts a DIFFERENT topic/task.
+
+Answer ONLY "NEW" or "SAME".
+
+Rules:
+- "NEW" = the new message is about a completely different subject, project, or task
+- "SAME" = the new message continues, follows up on, or is closely related to the current topic
+- Follow-up questions, clarifications, refinements, bug fixes, or next steps on the same task = SAME
+- Greetings or meta-questions like "你好" or "谢谢" without new substance = SAME
+- A clearly unrelated request (e.g., current topic is deployment, new message asks about cooking) = NEW
+
+Output exactly one word: NEW or SAME`;
+
+export async function judgeNewTopicBedrock(
+  currentContext: string,
+  newMessage: string,
+  cfg: SummarizerConfig,
+  log: Logger,
+): Promise<boolean> {
+  const model = cfg.model ?? "anthropic.claude-3-haiku-20240307-v1:0";
+  const endpoint = cfg.endpoint;
+  if (!endpoint) {
+    throw new Error("Bedrock topic-judge requires 'endpoint'");
+  }
+
+  const url = `${endpoint}/model/${model}/converse`;
+  const headers: Record<string, string> = {
+    "Content-Type": "application/json",
+    ...cfg.headers,
+  };
+
+  const userContent = `CURRENT CONVERSATION SUMMARY:\n${currentContext}\n\nNEW USER MESSAGE:\n${newMessage}`;
+
+  const resp = await fetch(url, {
+    method: "POST",
+    headers,
+    body: JSON.stringify({
+      system: [{ text: TOPIC_JUDGE_PROMPT }],
+      messages: [{ role: "user", content: [{ text: userContent }] }],
+      inferenceConfig: { temperature: 0, maxTokens: 10 },
+    }),
+    signal: AbortSignal.timeout(cfg.timeoutMs ?? 15_000),
+  });
+
+  if (!resp.ok) {
+    const body = await resp.text();
+    throw new Error(`Bedrock topic-judge failed (${resp.status}): ${body}`);
+  }
+
+  const json = (await resp.json()) as { output: { message: { content: Array<{ text: string }> } } };
+  const answer = json.output?.message?.content?.[0]?.text?.trim().toUpperCase() ?? "";
+  log.debug(`Topic judge result: "${answer}"`);
+  return answer.startsWith("NEW");
+}
+
+const FILTER_RELEVANT_PROMPT = `You are a memory relevance judge. Given a user's QUERY and a list of CANDIDATE memory summaries, do two things:
+
+1. Select ALL candidates that could be useful for answering the query. When in doubt, INCLUDE the candidate.
+   - For questions about lists, history, or "what/where/who" across multiple items (e.g. "which companies did I work at"), include ALL matching items — do NOT stop at the first match.
+   - For factual lookups (e.g. "what is the SSH port"), a single direct answer is enough.
+2. Judge whether the selected memories are SUFFICIENT to fully answer the query WITHOUT fetching additional context.
+
+IMPORTANT for "sufficient" judgment:
+- sufficient=true ONLY when the memories contain a concrete ANSWER, fact, decision, or actionable information that directly addresses the query.
+- sufficient=false when:
+  - The memories only repeat the same question the user asked before (echo, not answer).
+  - The memories show related topics but lack the specific detail needed.
+  - The memories contain partial information that would benefit from full task context, timeline, or related skills.
+
+Output a JSON object with exactly two fields:
+{"relevant":[1,3,5],"sufficient":true}
+
+- "relevant": array of candidate numbers that are useful. Empty array [] if none are relevant.
+- "sufficient": true ONLY if the memories contain a direct answer; false otherwise.
+
+Output ONLY the JSON object, nothing else.`;
+
+import type { FilterResult } from "./openai";
+export type { FilterResult } from "./openai";
+
+export async function filterRelevantBedrock(
+  query: string,
+  candidates: Array<{ index: number; summary: string; role: string }>,
+  cfg: SummarizerConfig,
+  log: Logger,
+): Promise<FilterResult> {
+  const model = cfg.model ?? "anthropic.claude-3-haiku-20240307-v1:0";
+  const endpoint = cfg.endpoint;
+  if (!endpoint) {
+    throw new Error("Bedrock filter-relevant requires 'endpoint'");
+  }
+
+  const url = `${endpoint}/model/${model}/converse`;
+  const headers: Record<string, string> = {
+    "Content-Type": "application/json",
+    ...cfg.headers,
+  };
+
+  const candidateText = candidates
+    .map((c) => `${c.index}. [${c.role}] ${c.summary}`)
+    .join("\n");
+
+  const resp = await fetch(url, {
+    method: "POST",
+    headers,
+    body: JSON.stringify({
+      system: [{ text: FILTER_RELEVANT_PROMPT }],
+      messages: [{ role: "user", content: [{ text: `QUERY: ${query}\n\nCANDIDATES:\n${candidateText}` }] }],
+      inferenceConfig: { temperature: 0, maxTokens: 200 },
+    }),
+    signal: AbortSignal.timeout(cfg.timeoutMs ?? 15_000),
+  });
+
+  if (!resp.ok) {
+    const body = await resp.text();
+    throw new Error(`Bedrock filter-relevant failed (${resp.status}): ${body}`);
+  }
+
+  const json = (await resp.json()) as { output: { message: { content: Array<{ text: string }> } } };
+  const raw = json.output?.message?.content?.[0]?.text?.trim() ?? "{}";
+  return parseFilterResult(raw, log);
+}
+
+function parseFilterResult(raw: string, log: Logger): FilterResult {
+  try {
+    const match = raw.match(/\{[\s\S]*\}/);
+    if (match) {
+      const obj = JSON.parse(match[0]);
+      if (obj && Array.isArray(obj.relevant)) {
+        return {
+          relevant: obj.relevant.filter((n: any) => typeof n === "number"),
+          sufficient: obj.sufficient === true,
+        };
+      }
+    }
+  } catch {}
+  log.warn(`filterRelevant: failed to parse LLM output: "${raw}", fallback to all+insufficient`);
+  return { relevant: [], sufficient: false };
+}
 
-/**
- * AWS Bedrock Converse API adapter.
- * Expects cfg.endpoint to be the full Bedrock invoke URL and
- * authentication handled via AWS SDK credential chain (env vars / IAM role).
- */
 export async function summarizeBedrock(
   text: string,
   cfg: SummarizerConfig,
@@ -48,3 +257,44 @@ export async function summarizeBedrock(
   };
   return json.output?.message?.content?.[0]?.text?.trim() ?? "";
 }
+
+// ─── Smart Dedup ───
+
+import { DEDUP_JUDGE_PROMPT, parseDedupResult } from "./openai";
+import type { DedupResult } from "./openai";
+export type { DedupResult } from "./openai";
+
+export async function judgeDedupBedrock(
+  newSummary: string,
+  candidates: Array<{ index: number; summary: string; chunkId: string }>,
+  cfg: SummarizerConfig,
+  log: Logger,
+): Promise<DedupResult> {
+  const model = cfg.model ?? "anthropic.claude-3-haiku-20240307-v1:0";
+  const endpoint = cfg.endpoint;
+  if (!endpoint) throw new Error("Bedrock dedup-judge requires 'endpoint'");
+
+  const url = `${endpoint}/model/${model}/converse`;
+  const headers: Record<string, string> = { "Content-Type": "application/json", ...cfg.headers };
+  const candidateText = candidates.map((c) => `${c.index}. ${c.summary}`).join("\n");
+
+  const resp = await fetch(url, {
+    method: "POST",
+    headers,
+    body: JSON.stringify({
+      system: [{ text: DEDUP_JUDGE_PROMPT }],
+      messages: [{ role: "user", content: [{ text: `NEW MEMORY:\n${newSummary}\n\nEXISTING MEMORIES:\n${candidateText}` }] }],
+      inferenceConfig: { temperature: 0, maxTokens: 300 },
+    }),
+    signal: AbortSignal.timeout(cfg.timeoutMs ?? 15_000),
+  });
+
+  if (!resp.ok) {
+    const body = await resp.text();
+    throw new Error(`Bedrock dedup-judge failed (${resp.status}): ${body}`);
+  }
+
+  const json = (await resp.json()) as { output: { message: { content: Array<{ text: string }> } } };
+  const raw = json.output?.message?.content?.[0]?.text?.trim() ?? "{}";
+  return parseDedupResult(raw, log);
+}

package/src/ingest/providers/gemini.ts

@@ -1,6 +1,217 @@
 import type { SummarizerConfig, Logger } from "../../types";
 
-const SYSTEM_PROMPT = `Summarize the text in ONE concise sentence (max 60 tokens). Preserve exact names, commands, error codes. No bullet points, no preamble — output only the sentence.`;
+const SYSTEM_PROMPT = `Summarize the text in ONE concise sentence (max 120 characters). IMPORTANT: Use the SAME language as the input text — if the input is Chinese, write Chinese; if English, write English. Preserve exact names, commands, error codes. No bullet points, no preamble — output only the sentence.`;
+
+const TASK_SUMMARY_PROMPT = `You create a DETAILED task summary from a multi-turn conversation. This summary will be the ONLY record of this conversation, so it must preserve ALL important information.
+
+CRITICAL LANGUAGE RULE: You MUST write in the SAME language as the user's messages. Chinese input → Chinese output. English input → English output. NEVER mix languages.
+
+Output EXACTLY this structure:
+
+📌 Title
+A short, descriptive title (10-30 characters). Like a chat group name.
+
+🎯 Goal
+One sentence: what the user wanted to accomplish.
+
+📋 Key Steps
+- Describe each meaningful step in detail
+- Include the ACTUAL content produced: code snippets, commands, config blocks, formulas, key paragraphs
+- For code: include the function signature and core logic (up to ~30 lines per block), use fenced code blocks
+- For configs: include the actual config values and structure
+- For lists/instructions: include the actual items, not just "provided a list"
+- Merge only truly trivial back-and-forth (like "ok" / "sure")
+- Do NOT over-summarize: "provided a function" is BAD; show the actual function
+
+✅ Result
+What was the final outcome? Include the final version of any code/config/content produced.
+
+💡 Key Details
+- Decisions made, trade-offs discussed, caveats noted, alternative approaches mentioned
+- Specific values: numbers, versions, thresholds, URLs, file paths, model names
+- Omit this section only if there truly are no noteworthy details
+
+RULES:
+- This summary is a KNOWLEDGE BASE ENTRY, not a brief note. Be thorough.
+- PRESERVE verbatim: code, commands, URLs, file paths, error messages, config values, version numbers, names, amounts
+- DISCARD only: greetings, filler, the assistant explaining what it will do before doing it
+- Replace secrets (API keys, tokens, passwords) with [REDACTED]
+- Target length: 30-50% of the original conversation length. Longer conversations need longer summaries.
+- Output summary only, no preamble.`;
+
+export async function summarizeTaskGemini(
+  text: string,
+  cfg: SummarizerConfig,
+  log: Logger,
+): Promise<string> {
+  const model = cfg.model ?? "gemini-1.5-flash";
+  const endpoint =
+    cfg.endpoint ??
+    `https://generativelanguage.googleapis.com/v1beta/models/${model}:generateContent`;
+
+  const url = `${endpoint}?key=${cfg.apiKey}`;
+  const headers: Record<string, string> = {
+    "Content-Type": "application/json",
+    ...cfg.headers,
+  };
+
+  const resp = await fetch(url, {
+    method: "POST",
+    headers,
+    body: JSON.stringify({
+      systemInstruction: { parts: [{ text: TASK_SUMMARY_PROMPT }] },
+      contents: [{ parts: [{ text }] }],
+      generationConfig: { temperature: cfg.temperature ?? 0.1, maxOutputTokens: 4096 },
+    }),
+    signal: AbortSignal.timeout(cfg.timeoutMs ?? 60_000),
+  });
+
+  if (!resp.ok) {
+    const body = await resp.text();
+    throw new Error(`Gemini task-summarize failed (${resp.status}): ${body}`);
+  }
+
+  const json = (await resp.json()) as { candidates: Array<{ content: { parts: Array<{ text: string }> } }> };
+  return json.candidates?.[0]?.content?.parts?.[0]?.text?.trim() ?? "";
+}
+
+const TOPIC_JUDGE_PROMPT = `You are a conversation topic boundary detector. Given a summary of the CURRENT conversation and a NEW user message, determine if the new message starts a DIFFERENT topic/task.
+
+Answer ONLY "NEW" or "SAME".
+
+Rules:
+- "NEW" = the new message is about a completely different subject, project, or task
+- "SAME" = the new message continues, follows up on, or is closely related to the current topic
+- Follow-up questions, clarifications, refinements, bug fixes, or next steps on the same task = SAME
+- Greetings or meta-questions like "你好" or "谢谢" without new substance = SAME
+- A clearly unrelated request (e.g., current topic is deployment, new message asks about cooking) = NEW
+
+Output exactly one word: NEW or SAME`;
+
+export async function judgeNewTopicGemini(
+  currentContext: string,
+  newMessage: string,
+  cfg: SummarizerConfig,
+  log: Logger,
+): Promise<boolean> {
+  const model = cfg.model ?? "gemini-1.5-flash";
+  const endpoint =
+    cfg.endpoint ??
+    `https://generativelanguage.googleapis.com/v1beta/models/${model}:generateContent`;
+
+  const url = `${endpoint}?key=${cfg.apiKey}`;
+  const headers: Record<string, string> = {
+    "Content-Type": "application/json",
+    ...cfg.headers,
+  };
+
+  const userContent = `CURRENT CONVERSATION SUMMARY:\n${currentContext}\n\nNEW USER MESSAGE:\n${newMessage}`;
+
+  const resp = await fetch(url, {
+    method: "POST",
+    headers,
+    body: JSON.stringify({
+      systemInstruction: { parts: [{ text: TOPIC_JUDGE_PROMPT }] },
+      contents: [{ parts: [{ text: userContent }] }],
+      generationConfig: { temperature: 0, maxOutputTokens: 10 },
+    }),
+    signal: AbortSignal.timeout(cfg.timeoutMs ?? 15_000),
+  });
+
+  if (!resp.ok) {
+    const body = await resp.text();
+    throw new Error(`Gemini topic-judge failed (${resp.status}): ${body}`);
+  }
+
+  const json = (await resp.json()) as { candidates: Array<{ content: { parts: Array<{ text: string }> } }> };
+  const answer = json.candidates?.[0]?.content?.parts?.[0]?.text?.trim().toUpperCase() ?? "";
+  log.debug(`Topic judge result: "${answer}"`);
+  return answer.startsWith("NEW");
+}
+
+const FILTER_RELEVANT_PROMPT = `You are a memory relevance judge. Given a user's QUERY and a list of CANDIDATE memory summaries, do two things:
+
+1. Select ALL candidates that could be useful for answering the query. When in doubt, INCLUDE the candidate.
+   - For questions about lists, history, or "what/where/who" across multiple items (e.g. "which companies did I work at"), include ALL matching items — do NOT stop at the first match.
+   - For factual lookups (e.g. "what is the SSH port"), a single direct answer is enough.
+2. Judge whether the selected memories are SUFFICIENT to fully answer the query WITHOUT fetching additional context.
+
+IMPORTANT for "sufficient" judgment:
+- sufficient=true ONLY when the memories contain a concrete ANSWER, fact, decision, or actionable information that directly addresses the query.
+- sufficient=false when:
+  - The memories only repeat the same question the user asked before (echo, not answer).
+  - The memories show related topics but lack the specific detail needed.
+  - The memories contain partial information that would benefit from full task context, timeline, or related skills.
+
+Output a JSON object with exactly two fields:
+{"relevant":[1,3,5],"sufficient":true}
+
+- "relevant": array of candidate numbers that are useful. Empty array [] if none are relevant.
+- "sufficient": true ONLY if the memories contain a direct answer; false otherwise.
+
+Output ONLY the JSON object, nothing else.`;
+
+import type { FilterResult } from "./openai";
+export type { FilterResult } from "./openai";
+
+export async function filterRelevantGemini(
+  query: string,
+  candidates: Array<{ index: number; summary: string; role: string }>,
+  cfg: SummarizerConfig,
+  log: Logger,
+): Promise<FilterResult> {
+  const model = cfg.model ?? "gemini-1.5-flash";
+  const endpoint =
+    cfg.endpoint ??
+    `https://generativelanguage.googleapis.com/v1beta/models/${model}:generateContent`;
+
+  const url = `${endpoint}?key=${cfg.apiKey}`;
+  const headers: Record<string, string> = {
+    "Content-Type": "application/json",
+    ...cfg.headers,
+  };
+
+  const candidateText = candidates
+    .map((c) => `${c.index}. [${c.role}] ${c.summary}`)
+    .join("\n");
+
+  const resp = await fetch(url, {
+    method: "POST",
+    headers,
+    body: JSON.stringify({
+      systemInstruction: { parts: [{ text: FILTER_RELEVANT_PROMPT }] },
+      contents: [{ parts: [{ text: `QUERY: ${query}\n\nCANDIDATES:\n${candidateText}` }] }],
+      generationConfig: { temperature: 0, maxOutputTokens: 200 },
+    }),
+    signal: AbortSignal.timeout(cfg.timeoutMs ?? 15_000),
+  });
+
+  if (!resp.ok) {
+    const body = await resp.text();
+    throw new Error(`Gemini filter-relevant failed (${resp.status}): ${body}`);
+  }
+
+  const json = (await resp.json()) as { candidates: Array<{ content: { parts: Array<{ text: string }> } }> };
+  const raw = json.candidates?.[0]?.content?.parts?.[0]?.text?.trim() ?? "{}";
+  return parseFilterResult(raw, log);
+}
+
+function parseFilterResult(raw: string, log: Logger): FilterResult {
+  try {
+    const match = raw.match(/\{[\s\S]*\}/);
+    if (match) {
+      const obj = JSON.parse(match[0]);
+      if (obj && Array.isArray(obj.relevant)) {
+        return {
+          relevant: obj.relevant.filter((n: any) => typeof n === "number"),
+          sufficient: obj.sufficient === true,
+        };
+      }
+    }
+  } catch {}
+  log.warn(`filterRelevant: failed to parse LLM output: "${raw}", fallback to all+insufficient`);
+  return { relevant: [], sufficient: false };
+}
 
 export async function summarizeGemini(
   text: string,
@@ -39,3 +250,43 @@ export async function summarizeGemini(
   };
   return json.candidates?.[0]?.content?.parts?.[0]?.text?.trim() ?? "";
 }
+
+// ─── Smart Dedup ───
+
+import { DEDUP_JUDGE_PROMPT, parseDedupResult } from "./openai";
+import type { DedupResult } from "./openai";
+export type { DedupResult } from "./openai";
+
+export async function judgeDedupGemini(
+  newSummary: string,
+  candidates: Array<{ index: number; summary: string; chunkId: string }>,
+  cfg: SummarizerConfig,
+  log: Logger,
+): Promise<DedupResult> {
+  const model = cfg.model ?? "gemini-1.5-flash";
+  const endpoint = cfg.endpoint ?? `https://generativelanguage.googleapis.com/v1beta/models/${model}:generateContent`;
+  const url = `${endpoint}?key=${cfg.apiKey}`;
+  const headers: Record<string, string> = { "Content-Type": "application/json", ...cfg.headers };
+
+  const candidateText = candidates.map((c) => `${c.index}. ${c.summary}`).join("\n");
+
+  const resp = await fetch(url, {
+    method: "POST",
+    headers,
+    body: JSON.stringify({
+      systemInstruction: { parts: [{ text: DEDUP_JUDGE_PROMPT }] },
+      contents: [{ parts: [{ text: `NEW MEMORY:\n${newSummary}\n\nEXISTING MEMORIES:\n${candidateText}` }] }],
+      generationConfig: { temperature: 0, maxOutputTokens: 300 },
+    }),
+    signal: AbortSignal.timeout(cfg.timeoutMs ?? 15_000),
+  });
+
+  if (!resp.ok) {
+    const body = await resp.text();
+    throw new Error(`Gemini dedup-judge failed (${resp.status}): ${body}`);
+  }
+
+  const json = (await resp.json()) as { candidates: Array<{ content: { parts: Array<{ text: string }> } }> };
+  const raw = json.candidates?.[0]?.content?.parts?.[0]?.text?.trim() ?? "{}";
+  return parseDedupResult(raw, log);
+}