`query_engine.py`

This module defines the core logic for extracting answers from documents using a Retrieval-Augmented Generation (RAG) pipeline powered by Groq LLM.

Responsibilities

Embed and semantically search chunks for each document via FAISS
Generate answers from top-k retrieved chunks using Groq's LLM
Return detailed, clause-preserving responses along with paragraph-level citations

Core Function

`extract_answers_from_docs(session_id: str, user_query: str, doc_ids: List[str]) -> List[Dict[str, str]]`

Given the session_id, a user question and a list of document IDs:

Retrieves top-k chunks per document using FAISS
Constructs a detailed system prompt including:
- Chunk texts
- Paragraph references
- The user’s question
Sends the prompt to Groq LLM
Parses the LLM's response (JSON only)
Returns answers and citations for each document

Prompt Design

The prompt sent to Groq instructs the LLM to:

Extract full, unmodified clauses, sections, and references
Avoid paraphrasing or summarizing legal/technical terms
Return a strict JSON format:
```
{"answer":"...","citation":"Para X"}
```

Citation Extraction
Each paragraph is referenced as [Para X] in the search prompt

The LLM is instructed to cite the specific paragraph(s)

Invalid responses or “no relevant information” are gracefully handled

Dependencies

Groq – LLM client (llama-3.3-70b-versatile used)
search_top_k_chunks() from vector_store.py – for semantic chunk retrieval
config.GROQ_API_KEY – pulled from .env
logger from logger.py for structured logging
re, json, tabulate – for response validation and formatting

Notes

Responses are forced into valid JSON with RegEx pre-checking
Empty or badly formatted responses are skipped with warnings
Errors are logged and do not crash the pipeline

Related Modules

vector_store.py – Provides search_top_k_chunks

pipeline_routes.py – Calls this function during analysis

theme_identifier.py – Uses results from this module to extract themes

FilesExpand file tree

query_engine.MD

Latest commit

History