5 files changed, 303 insertions, 0 deletions

diff --git a/documents/doc1.txt b/documents/doc1.txt
new file mode 100644
index 0000000..56f5cd5
--- /dev/null
+++ b/documents/doc1.txt
@@ -0,0 +1,3 @@
+horse number 77331893112373437 jumped over a greyhound
+strawberries are red in colour
+mcdonalds serve fast food
diff --git a/documents/doc2.txt b/documents/doc2.txt
new file mode 100644
index 0000000..7d5a79d
--- /dev/null
+++ b/documents/doc2.txt
@@ -0,0 +1,3 @@
+john built a building named "wonkystairs" in 2002
+one of the locations of the september 11 attacks is new york
+"The Straits Times" is a newspaper from singapore
diff --git a/readme b/readme
new file mode 100644
index 0000000..8e37e85
--- /dev/null
+++ b/readme
@@ -0,0 +1,16 @@
+#Installation: (pytorch higher version should work as well, the gpu i have is a bit old)
+python3 -m venv venv && source venv/bin/activate
+pip install -r requirements.txt --extra-index-url https://download.pytorch.org/whl/cu121
+
+#Usage:
+./search.py "your research question"
+./search.py --test  # sanity check to test llm workability
+./search.py "what does mcdonalds serve?"
+./search.py "is new york one of the locations of the sept 11 attacks?"
+./search.py "strawberries. what colour are they?"
+
+#Documentation:
+This program uses two research tools:
+        - local document(in ./documents) search using an embedding model + ChromaDB for semantic retrieval
+        - web search using duckduckgo
+A local llm orchestrates the research loop, deciding when enough information has been gathered, then provides a final concise summary.
diff --git a/requirements.txt b/requirements.txt
new file mode 100644
index 0000000..d003057
--- /dev/null
+++ b/requirements.txt
@@ -0,0 +1,9 @@
+torch==2.5.1+cu121
+torchvision==0.20.1+cu121
+transformers
+accelerate
+sentence-transformers
+chromadb
+httpx
+beautifulsoup4
+ddgs
diff --git a/search.py b/search.py
new file mode 100755
index 0000000..c370ffe
--- /dev/null
+++ b/search.py
@@ -0,0 +1,272 @@
+#!/usr/bin/env python3
+"""
+multi-agent deep research thingy
+"""
+
+
+import os, re, sys
+from dataclasses import dataclass, field
+from pathlib import Path
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from sentence_transformers import SentenceTransformer
+import chromadb
+import httpx
+from bs4 import BeautifulSoup
+from ddgs import DDGS
+
+# ==================== CONFIG ====================
+@dataclass
+class Config:
+    model_name: str = "Qwen/Qwen2.5-1.5B-Instruct"
+    embedding_model: str = "all-MiniLM-L6-v2"
+    device: str = field(default_factory=lambda: "cuda" if torch.cuda.is_available() else "cpu")
+    docs_dir: str = field(default_factory=lambda: os.getenv("DOCS_DIR", "./documents"))
+    max_critique_rounds: int = 3
+
+CFG = Config()
+log = lambda tag, msg: print(f"[{tag}] {msg}")
+
+# ==================== MODEL init ====================
+class _Models:
+    def __init__(self):
+        self._ready = False
+        self.embedder = self.tokenizer = self.llm = self.collection = None
+
+    def _init(self):
+        if self._ready: return
+        log("init", f"Device: {CFG.device}")
+        log("init", f"Loading embedder ({CFG.embedding_model})...")
+        self.embedder = SentenceTransformer(CFG.embedding_model, device=CFG.device)
+        log("init", f"Loading LLM ({CFG.model_name})...")
+        self.tokenizer = AutoTokenizer.from_pretrained(CFG.model_name)
+        self.llm = AutoModelForCausalLM.from_pretrained(
+            CFG.model_name, dtype=torch.float16, device_map="auto"
+        )
+        client = chromadb.Client()
+        self.collection = client.get_or_create_collection("research_docs", metadata={"hnsw:space": "cosine"})
+        log("init", "Ready.")
+        self._ready = True
+
+    def embed(self, text: str) -> list[float]:
+        self._init()
+        return self.embedder.encode(text).tolist()
+
+    def generate(self, task: str, instructions: str = "", max_tokens: int = 512) -> str:
+        self._init()
+        msgs = ([{"role": "system", "content": instructions}] if instructions else []) + [{"role": "user", "content": task}]
+        text = self.tokenizer.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
+        inputs = self.tokenizer(text, return_tensors="pt").to(self.llm.device)
+        with torch.no_grad():
+            out = self.llm.generate(**inputs, max_new_tokens=max_tokens, temperature=0.7,
+                                     do_sample=True, pad_token_id=self.tokenizer.eos_token_id)
+        return self.tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True).strip()
+
+M = _Models()
+
+# ==================== MEMORY ====================
+@dataclass
+class Memory:
+    findings: list = field(default_factory=list)
+
+    def save(self, source: str, query: str, content: str) -> str:
+        summary = (content[:300].replace("\n", " ").strip() + "...") if len(content) > 300 else content
+        self.findings.append({"source": source, "query": query, "summary": summary})
+        return summary
+
+    def by_source(self, src: str) -> list[dict]:
+        return [f for f in self.findings if f["source"] == src]
+
+    def all_summaries(self) -> str:
+        return "\n".join(f"- [{f['source']}] {f['query']}: {f['summary']}" for f in self.findings)
+
+# ==================== TOOLS ====================
+def web_search(query: str, max_results: int = 5) -> list[dict]:
+    try:
+        log("duck", f"Searching: {query}")
+        with DDGS() as ddgs:
+            raw = list(ddgs.text(query, max_results=max_results))
+        results = [{"title": r.get("title", ""), "snippet": r.get("body", ""), "url": r.get("href", "")} for r in raw]
+        for item in results[:2]:
+            item["content"] = fetch_url(item["url"])
+        return results
+    except Exception as e:
+        log("duck", f"Error: {e}")
+        return []
+
+def fetch_url(url: str, max_chars: int = 3000) -> str:
+    try:
+        r = httpx.get(url, timeout=2, follow_redirects=True,
+                      headers={"User-Agent": "Mozilla/5.0 (compatible; ResearchBot/1.0)"})
+        soup = BeautifulSoup(r.text, "html.parser")
+        for tag in soup(["script", "style", "nav", "header", "footer"]): tag.decompose()
+        return soup.get_text(separator="\n", strip=True)[:max_chars]
+    except Exception as e:
+        log("fetch", f"Failed: {e}")
+        return ""
+
+def doc_search(query: str, n_results: int = 5) -> list[dict]:
+    if M.collection.count() == 0: return []
+    results = M.collection.query(query_embeddings=[M.embed(query)], n_results=n_results, include=["documents", "distances"])
+    if not results["documents"] or not results["documents"][0]: return []
+    docs = results["documents"][0]
+    dists = results.get("distances", [[1.0] * len(docs)])[0]
+    return [{"content": d, "score": 1 - dist} for d, dist in zip(docs, dists)]
+
+def index_documents(docs_dir: str = None):
+    path = Path(docs_dir or CFG.docs_dir)
+    if not path.exists():
+        log("docs", f"Directory not found: {path}")
+        return
+    docs, ids = [], []
+    for f in path.rglob("*"):
+        if f.suffix not in {".txt"}: continue
+        try:
+            content = f.read_text()
+            log("docs", f"Loading: {f.name} ({len(content)} chars)")
+            docs.append(content)
+            ids.append(str(f))
+        except Exception as e:
+            log("docs", f"Failed to load {f}: {e}")
+    if docs:
+        embeddings = [M.embed(d) for d in docs]  # triggers _init()
+        M.collection.add(documents=docs, embeddings=embeddings, ids=ids)
+        log("docs", f"Indexed {len(docs)} chunks")
+
+# ==================== AGENTS ====================
+def parse_action(text: str) -> tuple[str | None, str]:
+    if m := re.search(r'\[\[(\w+):(.+?)\]\]', text, re.DOTALL): return m.group(1).upper(), m.group(2).strip()
+    if m := re.search(r'\[\[(\w+)\]\]', text): return m.group(1).upper(), ""
+    return None, ""
+
+def extract_findings(resp: str) -> str:
+    """Extract content from [[FINDINGS:...]] or return raw response."""
+    if m := re.search(r'\[\[FINDINGS:(.*?)\]\]', resp, re.DOTALL):
+        return m.group(1).strip()
+    return resp
+
+def agent(name: str, instructions: str, task: str, max_tokens: int = 512) -> str:
+    resp = M.generate(task, instructions=instructions, max_tokens=max_tokens)
+    log(name, resp[:1000] + ("..." if len(resp) > 1000 else ""))
+    return resp
+
+INSTRUCTIONS = {
+    "planner": """You are a research planner. Break the query into 3 subtopics MAX.
+Output EXACTLY: [[PLAN:\n- subtopic 1\n- subtopic 2\n]]\nKeep subtopics short (3-5 words). No explanations.""",
+
+    "researcher": """You are a research agent. Be CONCISE - max 2-3 sentences.
+Extract ANY facts from the documents that relate to the query.
+Output format: [[FINDINGS:\nThe relevant facts found.\n]]""",
+
+    "critic": """You are a research critic. Review findings for completeness and accuracy.
+If sufficient: [[SATISFIED]]
+If gaps exist: [[ISSUES:what specific information is missing]]
+Be concise and specific.""",
+
+    "writer": "You are a research writer. Be CONCISE and DIRECT. No fluff, no hedging. Just state the facts."
+}
+
+def plan(mem: Memory, query: str) -> list[str]:
+    resp = agent("planner", INSTRUCTIONS["planner"], f"Research query: {query}")
+    if m := re.search(r'\[\[PLAN:(.*?)\]\]', resp, re.DOTALL):
+        subtopics = [l.strip().lstrip("-").strip() for l in m.group(1).strip().split("\n")]
+        subtopics = [s for s in subtopics if len(s) > 3]
+        if subtopics:
+            mem.save("planner", query, "\n".join(subtopics))
+            return subtopics
+    return [query]
+
+def do_research(mem: Memory, query: str, source: str = "web"):
+    log("research", f"Searching {source.upper()} for: {query}")
+    results = doc_search(query) if source == "local" else web_search(query)
+    if not results:
+        log("research", f"No {source} results")
+        return
+    log("research", f"Found {len(results)} {source} results")
+    if source == "local":
+        content = "\n".join(f"[{i}] (sim: {r['score']:.2f})\n{r['content'][:1000]}" for i, r in enumerate(results, 1))
+    else:
+        content = "\n".join(f"[{i}] {r['title']}\n{r['url']}\n{r.get('content', r.get('snippet', ''))[:1000]}" for i, r in enumerate(results, 1))
+    prompt = f"Research query: {query}\n\nResults:\n{content[:3000]}\n\nExtract key findings."
+    findings = extract_findings(agent("research", INSTRUCTIONS["researcher"], prompt))
+    mem.save(source, query, findings)
+
+def critique(mem: Memory, query: str) -> tuple[bool, str]:
+    prompt = f"Original query: {query}\n\nResearch so far:\n{mem.all_summaries()}\n\nIs this sufficient?"
+    resp = agent("critic", INSTRUCTIONS["critic"], prompt, max_tokens=200)
+    action, arg = parse_action(resp)
+    if action == "SATISFIED": return True, "Research approved"
+    if action == "ISSUES":
+        mem.save("critic", "gap identified", arg)
+        return False, arg
+    return True, "Assumed complete"
+
+def write(mem: Memory, query: str) -> str:
+    fmt = lambda t, s, e: f"## {t}\n" + ("\n".join(f"- {f['summary']}" for f in mem.by_source(s)) or e)
+    sections = [
+        fmt("LOCAL DOCUMENTS", "local", "No relevant local documents."),
+        fmt("WEB SEARCH", "web", "No relevant web results.")
+    ]
+    prompt = f"Query: {query}\n\nFindings:\n{mem.all_summaries()}\n\nWrite a 2-3 sentence answer."
+    sections.append(f"## ANSWER\n{agent('writer', INSTRUCTIONS['writer'], prompt, 150)}")
+    return "\n\n".join(sections)
+
+# ==================== ORCHESTRATOR ====================
+def research(query: str, verbose: bool = True) -> dict:
+    vlog = (lambda phase, msg: print(f"\n[Phase {phase}] {msg}")) if verbose else (lambda *_: None)
+    if verbose: print(f"\n{'='*60}\nRESEARCH: {query}\n{'='*60}\n")
+
+    mem = Memory()
+    index_documents()
+
+    vlog(1, "Searching local documents...")
+    do_research(mem, query, "local")
+
+    vlog(2, "Planning web research...")
+    subtopics = plan(mem, query)
+    if verbose: print(f"Subtopics: {subtopics}\n")
+
+    vlog(3, "Web research...")
+    for topic in subtopics:
+        if verbose: print(f"\n--- Web: {topic} ---")
+        do_research(mem, topic, "web")
+
+    vlog(4, "Critique loop...")
+    for rnd in range(CFG.max_critique_rounds):
+        if verbose: print(f"\n--- Critique round {rnd+1} ---")
+        ok, feedback = critique(mem, query)
+        if ok:
+            if verbose: print("Critic satisfied")
+            break
+        if verbose: print(f"Gap: {feedback[:100]}...")
+        do_research(mem, feedback, "web")
+
+    vlog(5, "Writing synthesis...")
+    return {"query": query, "subtopics": subtopics, "answer": write(mem, query)}
+
+# ==================== CLI ====================
+def test_model():
+    log("test", "Loading model and asking: 'What is an apple?'")
+    resp = M.generate("What is an apple? Answer in 2-3 sentences.", max_tokens=100)
+    print(f"[response] {resp}\n[test] Done.")
+
+if __name__ == "__main__":
+    if len(sys.argv) < 2:
+        print("""
+Installation: (pytorch higher version should work as well, the gpu i have is a bit old)
+    pip install -r requirements.txt --extra-index-url https://download.pytorch.org/whl/cu121
+
+Usage:
+./search.py "your research question"
+./search.py --test  # sanity check to test llm workability
+
+./search.py "what does mcdonalds serve?"
+./search.py "is new york one of the locations of the sept 11 attacks?"
+./search.py "strawberries. what colour are they?"
+""")
+        sys.exit(1)
+    if sys.argv[1] == "--test":
+        test_model()
+    else:
+        result = research(sys.argv[1])
+        print(f"\n{'='*60}\nFINAL ANSWER\n{'='*60}\n{result['answer']}")