Switch to Qwen-1.5-1.8B-Chat - verified multilingual model with good Indic support

Pass HF_TOKEN explicitly to model loading - more reliable in Docker

Switch to AI4Bharat IndicLLM - better support for 11 Indic languages

update max new token to 1024

Simplify validation - use general patterns, not hardcoded lists

Set generation_config on model only, not passed to pipeline - fixes duplicate arg error

Suppress max_new_tokens/max_length warning - cosmetic only

Create gen_config once, pass to pipeline once - clean implementation

Use GenerationConfig object passed to pipeline - proper format

Pass generation params directly to pipeline - no GenerationConfig object

Remove redundant config override - set generation_config once

Force override model.config max_length and max_new_tokens to fix warning

Use structured prompt format (CONTEXT/OBJECTIVE/STYLE/TONE/AUDIENCE/RESPONSE)

Simplify prompt and validation - robust creative output

Fix syntax error: missing quote in invalid_patterns

Use comprehensive creative writer prompt with strict output validation

Fix prompt leakage: simpler prompt format, stricter filtering for exact leakage patterns

Set max_length=None in GenerationConfig to override model defaults

Fix max_length conflict by explicitly setting max_length=None in pipeline

Add poem/story specifications: type, theme, length, style

Add output validation and strengthen prompt against meta-commentary

Add user-selected language to prompt for proper language response

Restructure prompt for direct creative output, improve meta-commentary filtering

Improve output cleaning: remove </s>, [INST], bracket numbers

Strengthen prompt: never refer to or explain context

Add output cleaning to filter informal/garbled text

Update prompt for creative writing, increase temperature to 0.9 and top_p to 0.92

Set generation_config on model directly to avoid duplicate param error

Use GenerationConfig to avoid parameter conflict warnings

Fix max_length conflict, set max_new_tokens to 1024

Increase max_tokens to 4096, send top 3 docs, truncate context to 800 chars

Increase max_new_tokens to 1024 for longer responses

Update prompt for Hindi literature expertise and multilingual support

Fix prompt template and increase max_new_tokens to 512

Fix CPU inference: auto-detect GPU, use float16 on CPU

Use bitsandbytes 4-bit quantization instead of AirLLM (more stable)

Add airllm and optimum to requirements

Use AirLLM 4-bit quantization for Sarvam-1 (uses ~1.5GB RAM)

Switch to TinyLlama-1.1B with float16 for lower memory

Tighten prompt to reduce meta-commentary

Improve system prompt for better RAG responses

Add debug logging for prompt/response

Optimize generation params for HF free tier CPU

Remove local_files_only from pipeline - not valid for model_kwargs

Fix Sarvam-1 model loading: enable download and use correct dtype parameter

Fix duplicate local_files_only keyword argument in Sarvam-1 initialization

Pre-download models in Dockerfile, use cache at runtime

merge local branch