File size: 9,480 Bytes
1075316
 
 
 
 
 
2833375
1075316
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a148d79
 
 
 
 
1075316
 
 
a148d79
1075316
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2833375
1075316
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2833375
1075316
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2833375
1075316
2833375
1075316
2833375
1075316
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2833375
1075316
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
import os
import re
import pickle
from pathlib import Path
from typing import List, Dict, Any

import streamlit as st
import numpy as np
import faiss
from sentence_transformers import SentenceTransformer

# ========= LLM backend config =========
USE_OPENAI = os.getenv("USE_OPENAI", "0") == "1"
GEMINI_MODEL = os.getenv("GEMINI_MODEL", "gemini-2.5-flash")
OPENAI_MODEL = os.getenv("OPENAI_MODEL", "gpt-4o-mini")

if USE_OPENAI:
    from openai import OpenAI
    OPENAI_API_KEY = os.getenv("OPENAI_API_KEY", "")
    if OPENAI_API_KEY:
        openai_client = OpenAI(api_key=OPENAI_API_KEY)
    else:
        openai_client = None
else:
    import google.generativeai as genai
    GOOGLE_API_KEY = os.getenv("GOOGLE_API_KEY", "")
    if GOOGLE_API_KEY:
        genai.configure(api_key=GOOGLE_API_KEY)

# ========= Page config =========
st.set_page_config(
    page_title="Halassa Lab Literature Chatbot",
    page_icon="🧠",
    layout="wide",
)

from pathlib import Path

BASE_DIR = Path(__file__).resolve().parent  # points to src/
DATA_DIR = BASE_DIR / "data"

VECTOR_PATH = DATA_DIR / "vector_store.index"
PKL_PATH = DATA_DIR / "data.pkl"


EMBED_MODEL_NAME = os.getenv("EMBED_MODEL_NAME", "BAAI/bge-large-en-v1.5")
TOP_K = int(os.getenv("TOP_K", "5"))
MAX_CONTEXT_CHARS = int(os.getenv("MAX_CONTEXT_CHARS", "12000"))
SUGGESTED_Q = int(os.getenv("SUGGESTED_Q", "4"))

# ========= Helpers =========
def load_index_and_data():
    if not VECTOR_PATH.exists() or not PKL_PATH.exists():
        st.error(f"Missing index or data:\n- {VECTOR_PATH}\n- {PKL_PATH}")
        st.stop()

    index = faiss.read_index(str(VECTOR_PATH))
    with open(PKL_PATH, "rb") as f:
        stored = pickle.load(f)

    texts = stored.get("texts", [])
    sources = stored.get("sources", [])
    meta = stored.get("meta", [None] * len(texts))

    if len(texts) == 0 or len(texts) != len(sources):
        st.error("data.pkl must contain 'texts' and 'sources' of equal length.")
        st.stop()

    return index, texts, sources, meta

@st.cache_resource(show_spinner=False)
def get_embedder():
    return SentenceTransformer(EMBED_MODEL_NAME)

def encode_query(query: str, embedder) -> np.ndarray:
    vec = embedder.encode([query])
    return vec.astype(np.float32)

def retrieve(query: str, index, texts, sources, meta, k=TOP_K):
    embedder = get_embedder()
    qvec = encode_query(query, embedder)
    D, I = index.search(qvec, k)
    results = []
    for rank, idx in enumerate(I[0].tolist()):
        if 0 <= idx < len(texts):
            results.append({
                "rank": rank + 1,
                "text": texts[idx],
                "source": sources[idx],
                "meta": meta[idx] if meta and idx < len(meta) else None
            })
    return results

def build_context(retrieved: List[Dict[str, Any]]) -> str:
    parts, total = [], 0
    for r in retrieved:
        src = r["source"]
        txt = r["text"].strip()
        chunk = f"Source: {src}\nContent: {txt}\n"
        if total + len(chunk) > MAX_CONTEXT_CHARS:
            break
        parts.append(chunk)
        total += len(chunk)
    return "\n---\n".join(parts)

def call_llm(system_prompt: str, user_prompt: str) -> str:
    # OpenAI path
    if USE_OPENAI and os.getenv("OPENAI_API_KEY") and openai_client:
        resp = openai_client.chat.completions.create(
            model=OPENAI_MODEL,
            messages=[
                {"role": "system", "content": system_prompt},
                {"role": "user", "content": user_prompt},
            ],
            temperature=0.2,
        )
        return resp.choices[0].message.content

    # Gemini path
    if not USE_OPENAI and os.getenv("GOOGLE_API_KEY"):
        model = genai.GenerativeModel(GEMINI_MODEL)
        resp = model.generate_content(system_prompt + "\n\n" + user_prompt)
        return resp.text

    # Fallback (no key) for UI testing
    return "(LLM disabled) " + user_prompt[:800]

def highlight_terms(text: str, query: str) -> str:
    # lightweight term highlighter
    import re
    terms = [t for t in re.split(r"\W+", query) if len(t) >= 3]
    out = text
    for t in set(terms):
        out = re.sub(rf"({re.escape(t)})", r"<mark>\1</mark>", out, flags=re.IGNORECASE)
    return out

def suggest_questions(last_answer: str, k=SUGGESTED_Q) -> List[str]:
    prompt = f"""Generate {k} concise follow-up questions a user might ask next, given the expert answer below.
Each question should be short (max ~12 words) and deepen the discussion.

Answer only with a bulletless list, one question per line.

Expert answer:
{last_answer}
"""
    out = call_llm(
        system_prompt="You are a helpful assistant that proposes follow-up questions.",
        user_prompt=prompt,
    )
    qs = [re.sub(r"^[\-\*\d\.\)\s]+", "", q).strip() for q in out.splitlines() if q.strip()]
    return [q for q in qs if q][:k]

# ========= Load index/data =========
index, TEXTS, SOURCES, META = load_index_and_data()

# ========= Sidebar =========
with st.sidebar:
    st.title("⚙️ Settings")
    st.write("**Retrieval**")
    TOP_K = st.slider("Top-K passages", 3, 10, TOP_K)
    st.divider()
    st.write("**Models**")
    st.write(f"Embedding: `{EMBED_MODEL_NAME}`")
    st.write("LLM:", "OpenAI" if USE_OPENAI else "Gemini",
             f"({OPENAI_MODEL if USE_OPENAI else GEMINI_MODEL})")
    st.caption("Switch with env vars: USE_OPENAI, OPENAI_API_KEY, GOOGLE_API_KEY.")
    st.divider()
    st.write("**Files**")
    st.write(f"Index: `{VECTOR_PATH}`")
    st.write(f"Data : `{PKL_PATH}`")

# ========= Main Layout =========
st.title("Halassa Lab Onboarder 🧠📄")
st.caption("Ask questions; see exactly which passages were used.")

if "chat" not in st.session_state:
    st.session_state.chat = []  # list[dict]: {"role": "user"/"assistant", "content": str, "retrieved": list}
if "last_suggestions" not in st.session_state:
    st.session_state.last_suggestions = []

# Input row
with st.container():
    cols = st.columns([6, 1])
    with cols[0]:
        user_message = st.text_input(
            "Ask your question",
            "",
            placeholder="e.g., How does MD dopamine shape error-driven flexibility?",
        )
    with cols[1]:
        ask = st.button("Send", use_container_width=True)

def answer_query(query: str):
    retrieved = retrieve(query, index, TEXTS, SOURCES, META, k=TOP_K)
    context_str = build_context(retrieved)

    sys_prompt = (
        "You are an Expert scientist in the Halassa Lab at MIT, expert in computational neuroscience. "
        "Answer thoroughly and clearly. Synthesize from provided context; write in your own words. "
        "If you cite directly from a provided paper, add citations at the end as [filename - Page X]. "
        "If context is partial, add helpful background."
    )
    user_prompt = f"""Context:
---
{context_str}
---

User Question: {query}

Expert Answer:
"""
    answer = call_llm(sys_prompt, user_prompt)

    st.session_state.chat.append({"role": "user", "content": query})
    st.session_state.chat.append({"role": "assistant", "content": answer, "retrieved": retrieved})

    try:
        st.session_state.last_suggestions = suggest_questions(answer, k=SUGGESTED_Q)
    except Exception:
        st.session_state.last_suggestions = []

# Trigger on click or Enter
if ask and user_message.strip():
    answer_query(user_message.strip())
elif user_message.strip() and st.session_state.chat == []:
    # allow pressing Enter to submit first question
    answer_query(user_message.strip())

# Two-column layout
col_chat, col_docs = st.columns([2, 1], gap="large")

# Left: Chat
with col_chat:
    for turn in st.session_state.chat:
        if turn["role"] == "user":
            st.chat_message("user").markdown(turn["content"])
        else:
            st.chat_message("assistant").markdown(turn["content"])

    if st.session_state.last_suggestions:
        st.subheader("Try next:")
        sug_cols = st.columns(len(st.session_state.last_suggestions))
        for i, q in enumerate(st.session_state.last_suggestions):
            if sug_cols[i].button(q):
                answer_query(q)

# Right: Relevant chunks (no PDF viewer)
with col_docs:
    st.subheader("Relevant Sources")
    last_assistant = None
    for t in reversed(st.session_state.chat):
        if t.get("role") == "assistant" and "retrieved" in t:
            last_assistant = t
            break

    if not last_assistant:
        st.info("Ask a question to see relevant passages.")
    else:
        # Find preceding user query for highlighting
        query_text = ""
        for i in range(len(st.session_state.chat)-1, -1, -1):
            if st.session_state.chat[i]["role"] == "user":
                query_text = st.session_state.chat[i]["content"]
                break

        for r in last_assistant["retrieved"]:
            src = r["source"]
            with st.expander(f"#{r['rank']}  {src}"):
                html = highlight_terms(r["text"], query_text)
                st.markdown(html, unsafe_allow_html=True)

                # Small utility buttons
                st.download_button(
                    "Download chunk",
                    data=r["text"].encode("utf-8"),
                    file_name=f"chunk_{r['rank']}.txt",
                    use_container_width=True
                )

st.divider()
st.caption("Tip: Ensure your `sources` strings match your citation format (e.g., `paper.pdf - Page 12`) so your LLM’s citations are clean.")