feat: add a detailed roadmap section to the README outlining future features and capabilities.

Author: CyberGhost007

README.md

@@ -308,6 +308,180 @@ After 25 iterations, top-scoring nodes are deep-read using GPT-4o vision on actu
 
 ---
 
+## Roadmap
+
+- [ ] **Adaptive MCTS**
+  Dynamically adjust iterations based on confidence. Stop early when the tree converges, explore deeper when scores are ambiguous. Prune low-scoring branches instead of revisiting them.
+  ```
+  Current:  Always runs 25 iterations → wastes calls on easy queries
+  Adaptive: Easy query → 8 iterations (converged early, saved 68% cost)
+            Hard query → 40 iterations (low confidence, kept exploring)
+  ```
+
+- [ ] **Multi-Hop Reasoning Chains**
+  Chain MCTS searches where each hop informs the next. The tree remembers which branches were relevant across hops.
+  ```
+  User: "What was Q3 budget vs the Q1 projection?"
+
+  Hop 1: MCTS → finds "Q3 Financial Summary" (pp.12-15) → extracts $4.2M
+  Hop 2: MCTS → searches with context "find Q1 projection to compare with $4.2M"
+         → finds "Q1 Forecast" (pp.3-4) → extracts $3.8M projected
+  Answer: "Q3 actual ($4.2M) exceeded Q1 projection ($3.8M) by 10.5%"
+  ```
+
+- [ ] **Absence Detection**
+  Exhaustively explore the tree to prove something *doesn't* exist. Vector RAG can never confirm absence — MCTS can.
+  ```
+  User: "Does this contract have a non-compete clause?"
+
+  MCTS: Visited 47/47 sections across 25 iterations
+        Coverage: 100% of document tree explored
+        Matches: 0 sections with relevance > 0.3
+  Answer: "No non-compete clause found. Searched all 47 sections
+           with full coverage. The contract covers: IP assignment,
+           confidentiality, termination — but no non-compete."
+  ```
+
+- [ ] **Structural Queries**
+  Answer questions about document *structure*, not just content. MCTS navigates the actual hierarchy — flat embeddings can't.
+  ```
+  User: "What comes after the methodology section?"
+
+  Tree traversal:
+    Root
+    ├── 1. Introduction
+    ├── 2. Literature Review
+    ├── 3. Methodology        ← found
+    ├── 4. Results & Analysis  ← next sibling
+    └── 5. Conclusion
+
+  Answer: "Chapter 4: Results & Analysis (pp.18-31) follows
+           the Methodology section. It covers three subsections:
+           4.1 Quantitative Results, 4.2 Qualitative Findings..."
+  ```
+
+- [ ] **Confidence-Bounded Answers**
+  Return interpretable confidence metrics: nodes visited vs total, convergence status, and score distribution.
+  ```
+  User: "What were the key risks identified?"
+
+  Search stats:
+    Nodes visited:  38/47 (81% coverage)
+    Converged:      Yes (iteration 19 of 25)
+    Top matches:    3 nodes, all in Chapter 6
+    Score spread:   0.92, 0.89, 0.85 (tight cluster = high confidence)
+
+  Answer: "Key risks identified (94% confidence): ..."
+  ```
+
+- [ ] **Progressive Drill-Down**
+  Resume the MCTS tree across follow-up queries instead of starting fresh. The tree is a persistent search state.
+  ```
+  User: "Give me an overview of this report"
+  MCTS: Explores level 1 nodes → returns top-level summary
+        Tree state saved ✓
+
+  User: "Tell me more about the financial section"
+  MCTS: Resumes from saved tree → deepens into "Finance" subtree
+        Skips already-explored branches → finds subsections
+        Tree state updated ✓
+
+  User: "Specifically the Q3 projections"
+  MCTS: Resumes → drills into Finance > Q3 > Projections leaf
+        3 queries, 1 tree, zero redundant exploration
+  ```
+
+- [ ] **Streaming Responses (SSE)**
+  Real-time answer generation with live phase indicators instead of 5-10s blank screen.
+  ```
+  [0.0s] ● Routing to folder "Project Zenith" (confidence: 94%)
+  [0.8s] ● Phase 1: Selecting documents... budget.pdf ✓ proposal.pdf ✓
+  [2.1s] ● Phase 2: Searching budget.pdf — iteration 12/25
+  [3.4s] ● Phase 2: Searching proposal.pdf — iteration 18/25
+  [4.2s] ● Deep reading pages 5-8 of budget.pdf...
+  [5.8s] ● Generating answer...
+  [6.1s] The total project cost was $812,000, broken down as...
+  ```
+
+- [ ] **Comparative Cross-Document Analysis**
+  Run parallel tree searches with a shared objective across multiple documents, finding the same structural section in each.
+  ```
+  User: "How do these three vendor contracts differ on liability?"
+
+  Parallel MCTS:
+    Contract A → "Section 8: Liability" (pp.12-14) → cap at $500K
+    Contract B → "Section 6: Liability" (pp.9-10)  → uncapped
+    Contract C → "Section 9: Liability" (pp.15-17) → cap at $1M
+
+  Answer:
+  | Clause        | Vendor A   | Vendor B   | Vendor C   |
+  |---------------|------------|------------|------------|
+  | Liability cap | $500K      | Uncapped ⚠ | $1M        |
+  | Indemnity     | Mutual     | One-way    | Mutual     |
+  | Insurance req | $2M        | None ⚠     | $5M        |
+  ```
+
+- [ ] **Query Decomposition**
+  Automatically break complex queries into sub-queries, each targeting a specific subtree.
+  ```
+  User: "Compare the methodology, results, and cost of Paper A vs Paper B"
+
+  Decomposed into 6 parallel searches:
+    Paper A → "Methodology" subtree  ──┐
+    Paper A → "Results" subtree      ──┼── synthesize
+    Paper A → "Cost" subtree         ──┤   comparison
+    Paper B → "Methodology" subtree  ──┤
+    Paper B → "Results" subtree      ──┤
+    Paper B → "Cost" subtree         ──┘
+
+  Answer: Structured comparison across all three dimensions
+  ```
+
+- [ ] **Contradiction Detection**
+  Systematically compare claims across sections of a document. Search for the same topic in every branch, extract stated figures, flag inconsistencies.
+  ```
+  User: "Are there any contradictions in this report?"
+
+  MCTS scans all branches for overlapping claims:
+    ⚠ Revenue conflict:
+      Page 12 (Executive Summary): "Annual revenue: $4.2M"
+      Page 38 (Financial Details): "Total revenue: $3.8M"
+    ⚠ Timeline conflict:
+      Page 5 (Overview): "Project started March 2024"
+      Page 22 (Timeline): "Kickoff date: January 2024"
+    ✓ Headcount consistent: 47 employees (mentioned 3 times)
+  ```
+
+- [ ] **Multi-Provider LLM Support**
+  Swap between OpenAI, Claude, Gemini, and local models (Ollama).
+  ```python
+  # .env — just change the provider
+  LLM_PROVIDER=anthropic          # or: openai, google, ollama
+  SEARCH_MODEL=claude-haiku-4-5   # cheap model for MCTS
+  ANSWER_MODEL=claude-sonnet-4-6  # heavy model for answers
+
+  # Local mode — no API costs
+  LLM_PROVIDER=ollama
+  SEARCH_MODEL=llama3:8b
+  ANSWER_MODEL=llama3:70b
+  ```
+
+- [ ] **Persistent Chat History**
+  SQLite-backed conversations with named sessions, search over past Q&A, and session continuity across restarts.
+  ```
+  User: "Show my recent sessions"
+
+  Sessions:
+    1. "Q3 Budget Analysis"    — 12 messages, 3 docs  (2h ago)
+    2. "Contract Review"       — 8 messages, 5 docs   (yesterday)
+    3. "Research Paper Notes"  — 24 messages, 2 docs   (3 days ago)
+
+  User: "Resume session 2"
+  → Restores full chat history + MCTS tree states
+  ```
+
+---
+
 ## License
 
 [MIT](LICENSE)